Path: blob/master/section-2-data-science-and-ml-tools/introduction-to-scikit-learn.ipynb
874 views
View source code | Read notebook in online book format
A Quick Machine Learning Modelling Tutorial with Python and Scikit-Learn
This notebook goes through a range of common and useful featues of the Scikit-Learn library.
There's a bunch here but I'm calling it quick because of how vast the Scikit-Learn library is.
Covering everything requires a full-blown documentation, of which, if you ever get stuck, I'd highly recommend checking out.
What is Scikit-Learn (sklearn)?
Scikit-Learn, also referred to as sklearn, is an open-source Python machine learning library.
It's built on top on NumPy (Python library for numerical computing) and Matplotlib (Python library for data visualization).
Why Scikit-Learn?
Although the fields of data science and machine learning are vast, the main goal is finding patterns within data and then using those patterns to make predictions.
And there are certain categories which a majority of problems fall into.
If you're trying to create a machine learning model to predict whether an email is spam and or not spam, you're working on a classification problem (whether something is one thing or another).
If you're trying to create a machine learning model to predict the price of houses given their characteristics, you're working on a regression problem (predicting a number).
If you're trying to get a machine learning algorithm to group together similar samples (that you don't necessarily know which should go together), you're working on a clustering problem.
Once you know what kind of problem you're working on, there are also similar steps you'll take for each.
Steps like splitting the data into different sets, one for your machine learning algorithms to learn on (the training set) and another to test them on (the testing set).
Choosing a machine learning model and then evaluating whether or not your model has learned anything.
Scikit-Learn offers Python implementations for doing all of these kinds of tasks (from preparing data to modelling data). Saving you from having to build them from scratch.
What does this notebook cover?
The Scikit-Learn library is very capable. However, learning everything off by heart isn't necessary. Instead, this notebook focuses some of the main use cases of the library.
More specifically, we'll cover:
An end-to-end Scikit-Learn worfklow
Getting the data ready
Choosing the right maching learning estimator/aglorithm/model for your problem
Fitting your chosen machine learning model to data and using it to make a prediction
Evaluting a machine learning model
Improving predictions through experimentation (hyperparameter tuning)
Saving and loading a pretrained model
Putting it all together in a pipeline
Note: All of the steps in this notebook are focused on supervised learning (having data and labels). The other side of supervised learning is unsupervised learning (having data but no labels).
After going through it, you'll have the base knolwedge of Scikit-Learn you need to keep moving forward.
Where can I get help?
If you get stuck or think of something you'd like to do which this notebook doesn't cover, don't fear!
The recommended steps you take are:
Try it - Since Scikit-Learn has been designed with usability in mind, your first step should be to use what you know and try figure out the answer to your own question (getting it wrong is part of the process). If in doubt, run your code.
Press SHIFT+TAB - See you can the docstring of a function (information on what the function does) by pressing SHIFT + TAB inside it. Doing this is a good habit to develop. It'll improve your research skills and give you a better understanding of the library.
Search for it - If trying it on your own doesn't work, since someone else has probably tried to do something similar, try searching for your problem. You'll likely end up in 1 of 2 places:
Scikit-Learn documentation/user guide - the most extensive resource you'll find for Scikit-Learn information.
Stack Overflow - this is the developers Q&A hub, it's full of questions and answers of different problems across a wide range of software development topics and chances are, there's one related to your problem.
ChatGPT - ChatGPT is very good at explaining code, however, it can make mistakes. Best to verify the code it writes first before using it. Try asking "Can you explain the following code for me? {your code here}" and then continue with follow up questions from there. Avoid blindly trusting code you didn't write for yourself.
An example of searching for a Scikit-Learn solution might be:
"how to tune the hyperparameters of a sklearn model"
Searching this on Google leads to the Scikit-Learn documentation for the GridSearchCV function: http://scikit-learn.org/stable/modules/grid_search.html
The next steps here are to read through the documentation, check the examples and see if they line up to the problem you're trying to solve. If they do, rewrite the code to suit your needs, run it, and see what the outcomes are.
Ask for help - If you've been through the above 3 steps and you're still stuck, you might want to ask your question on Stack Overflow or in the ZTM Machine Learning and AI Discord channel. Be as specific as possible and provide details on what you've tried.
Remember, you don't have to learn all of the functions off by heart to begin with.
What's most important is continually asking yourself, "what am I trying to do with the data?".
Start by answering that question and then practicing finding the code which does it.
Let's get started.
First we'll import the libraries we've been using previously.
We'll also check the version of sklearn we've got.
0. An end-to-end Scikit-Learn workflow
Before we get in-depth, let's quickly check out what an end-to-end Scikit-Learn workflow might look like.
Once we've seen an end-to-end workflow, we'll dive into each step a little deeper.
Specifically, we'll get hands-on with the following steps:
Getting data ready (split into features and labels, prepare train and test steps)
Choosing a model for our problem
Fit the model to the data and use it to make a prediction
Evaluate the model
Experiment to improve
Save a model for someone else to use
Note: The following section is a bit information heavy but it is an end-to-end workflow. We'll go through it quite swiftly but we'll break it down more throughout the rest of the notebook. And since Scikit-Learn is such a vast library, capable of tackling many problems, the workflow we're using is only one example of how you can use it.
Random Forest Classifier Workflow for Classifying Heart Disease
1. Get the data ready
As an example dataset, we'll import heart-disease.csv.
This file contains anonymised patient medical records and whether or not they have heart disease or not (this is a classification problem since we're trying to predict whether something is one thing or another).
Here, each row is a different patient and all columns except target are different patient characteristics.
The target column indicates whether the patient has heart disease (target=1) or not (target=0), this is our "label" columnm, the variable we're going to try and predict.
The rest of the columns (often called features) are what we'll be using to predict the target value.
Note: It's a common custom to save features to a varialbe
Xand labels to a variabley. In practice, we'd like to use theX(features) to build a predictive algorithm to predict they(labels).
One of the most important practices in machine learning is to split datasets into training and test sets.
As in, a model will train on the training set to learn patterns and then those patterns can be evaluated on the test set.
Crucially, a model should never see testing data during training.
This is equivalent to a student studying course materials during the semester (training set) and then testing their abilities on the following exam (testing set).
Scikit-learn provides the sklearn.model_selection.train_test_split method to split datasets in training and test sets.
Note: A common practice to use an 80/20 or 70/30 or 75/25 split for training/testing data. There is also a third set, known as a validation set (e.g. 70/15/15 for training/validation/test) for hyperparamter tuning on but for now we'll focus on training and test sets.
2. Choose the model and hyperparameters
Choosing a model often depends on the type of problem you're working on.
For example, there are different models that Scikit-Learn recommends whether you're working on a classification or regression problem.
You can see a map breaking down the different kinds of model options and recommendations in the Scikit-Learn documentation.
Scikit-Learn refers to models as "estimators", however, they are often also referred to as model or clf (short for classifier).
A model's hyperparameters are settings you can change to adjust it for your problem, much like knobs on an oven you can tune to cook your favourite dish.
We can see the current hyperparameters of a model with the get_params() method.
We'll leave this as is for now, as Scikit-Learn models generally have good default settings.
3. Fit the model to the data and use it to make a prediction
Fitting a model a dataset involves passing it the data and asking it to figure out the patterns.
If there are labels (supervised learning), the model tries to work out the relationship between the data and the labels.
If there are no labels (unsupervised learning), the model tries to find patterns and group similar samples together.
Most Scikit-Learn models have the fit(X, y) method built-in, where the X parameter is the features and the y parameter is the labels.
In our case, we start by fitting a model on the training split (X_train, y_train).
Use the model to make a prediction
The whole point of training a machine learning model is to use it to make some kind of prediction in the future.
Once your model instance is trained, you can use the predict() method to predict a target value given a set of features.
In other words, use the model, along with some new, unseen and unlabelled data to predict the label.
Note: Data you predict on should be in the same shape and format as data you trained on.
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
Cell In[10], line 2
1 # This doesn't work... incorrect shapes
----> 2 y_label = clf.predict(np.array([0, 2, 3, 4]))
, in ForestClassifier.predict(self, X)
883 def predict(self, X):
884 """
885 Predict class for X.
886
(...)
902 The predicted classes.
903 """
--> 904 proba = self.predict_proba(X)
906 if self.n_outputs_ == 1:
907 return self.classes_.take(np.argmax(proba, axis=1), axis=0)
, in ForestClassifier.predict_proba(self, X)
944 check_is_fitted(self)
945 # Check data
--> 946 X = self._validate_X_predict(X)
948 # Assign chunk of trees to jobs
949 n_jobs, _, _ = _partition_estimators(self.n_estimators, self.n_jobs)
, in BaseForest._validate_X_predict(self, X)
638 else:
639 force_all_finite = True
--> 641 X = self._validate_data(
642 X,
643 dtype=DTYPE,
644 accept_sparse="csr",
645 reset=False,
646 force_all_finite=force_all_finite,
647 )
648 if issparse(X) and (X.indices.dtype != np.intc or X.indptr.dtype != np.intc):
649 raise ValueError("No support for np.int64 index based sparse matrices")
, in BaseEstimator._validate_data(self, X, y, reset, validate_separately, cast_to_ndarray, **check_params)
631 out = X, y
632 elif not no_val_X and no_val_y:
--> 633 out = check_array(X, input_name="X", **check_params)
634 elif no_val_X and not no_val_y:
635 out = _check_y(y, **check_params)
, in check_array(array, accept_sparse, accept_large_sparse, dtype, order, copy, force_writeable, force_all_finite, ensure_2d, allow_nd, ensure_min_samples, ensure_min_features, estimator, input_name)
1043 else:
1044 msg = (
1045 f"Expected 2D array, got 1D array instead:\narray={array}.\n"
1046 "Reshape your data either using array.reshape(-1, 1) if "
1047 "your data has a single feature or array.reshape(1, -1) "
1048 "if it contains a single sample."
1049 )
-> 1050 raise ValueError(msg)
1052 if dtype_numeric and hasattr(array.dtype, "kind") and array.dtype.kind in "USV":
1053 raise ValueError(
1054 "dtype='numeric' is not compatible with arrays of bytes/strings."
1055 "Convert your data to numeric values explicitly instead."
1056 )
ValueError: Expected 2D array, got 1D array instead:
array=[0. 2. 3. 4.].
Reshape your data either using array.reshape(-1, 1) if your data has a single feature or array.reshape(1, -1) if it contains a single sample.
Oh no!
We get a ValueError (mismatched shapes):
This happens because we're trying to make predictions on data that is in a different format to the data our model was trained on.
Since our model was trained on data from X_train, predictions should be made on data in the same format and shape as X_train.
Our goal in many machine learning problems is to use patterns learned from the training data to make predictions on the test data (or future unseen data).
4. Evaluate the model
Now we've made some predictions, we can start to use some more Scikit-Learn methods to figure out how good our model is.
Each model or estimator has a built-in score() method.
This method compares how well the model was able to learn the patterns between the features and labels.
The score() method for each model uses a standard evaluation metric to measure your model's results.
In the case of a classifier (our model), one of the most common evaluation metrics is accuracy (the fraction of correct predictions out of total predictions).
Let's check out our model's accuracy on the training set.
Woah! Looks like our model does pretty well on the training datset.
This is because it has a chance to see both data and labels.
How about the test dataset?
Hmm, looks like our model's accuracy is a bit less on the test dataset than the training dataset.
This is quite often the case, because remember, a model has never seen the testing examples before.
There are also a number of other evaluation methods we can use for our classification models.
All of the following classification metrics come from the sklearn.metrics module:
classification_report(y_true, y_true)- Builds a text report showing various classification metrics such as precision, recall and F1-score.confusion_matrix(y_true, y_pred)- Create a confusion matrix to compare predictions to truth labels.accuracy_score(y_true, y_pred)- Find the accuracy score (the default metric) for a classifier.
All metrics have the following in common: they compare a model's predictions (y_pred) to truth labels (y_true).
5. Experiment to improve
The first model you build is often referred to as a baseline (a baseline is often even simpler than the model we've used, a baseline could be "let's just by default predict the most common value and then try to improve").
Once you've got a baseline model, like we have here, it's important to remember, this is often not the final model you'll use.
The next step in the workflow is to try and improve upon your baseline model.
How?
With one of the most important mottos in machine learning...
Experiment, experiment, experiment!
Experiments can come in many different forms.
But let's break it into two.
From a model perspective.
From a data perspective.
From a model perspective may involve things such as using a more complex model or tuning your models hyperparameters.
From a data perspective may involve collecting more data or better quality data so your existing model has more of a chance to learn the patterns within.
If you're already working on an existing dataset, it's often easier try a series of model perspective experiments first and then turn to data perspective experiments if you aren't getting the results you're looking for.
One thing you should be aware of is if you're tuning a models hyperparameters in a series of experiments, your reuslts should always be cross-validated (we'll see this later on!).
Cross-validation is a way of making sure the results you're getting are consistent across your training and test datasets (because it uses multiple versions of training and test sets) rather than just luck because of the order the original training and test sets were created.
Try different hyperparameters.
All different parameters should be cross-validated.
Note: Beware of cross-validation for time series problems (as for time series, you don't want to mix samples from the future with samples from the past).
Different models you use will have different hyperparameters you can tune.
For the case of our model, the RandomForestClassifier(), we'll start trying different values for n_estimators (a measure for the number of trees in the random forest).
By default, n_estimators=100, so how about we try values from 100 to 200 and see what happens (generally more is better)?
The metrics above were measured on a single train and test split.
Let's use sklearn.model_selection.cross_val_score to measure the results across 5 different train and test sets.
We can achieve this by setting cross_val_score(X, y, cv=5).
Where X is the full feature set and y is the full label set and cv is the number of train and test splits cross_val_score will automatically create from the data (in our case, 5 different splits, this is known as 5-fold cross-validation).
Which model had the best cross-validation score?
This is usually a better indicator of a quality model than a single split accuracy score.
Rather than set up and track the results of these experiments manually, we can get Scikit-Learn to do the exploration for us.
Scikit-Learn's sklearn.model_selection.GridSearchCV is a way to search over a set of different hyperparameter values and automatically track which perform the best.
Let's test it!
We can extract the best model/estimator with the best_estimator_ attribute.
And now we've got the best cross-validated model, we can fit and score it on our original single train/test split of the data.
6. Save a model for someone else to use
When you've done a few experiments and you're happy with how your model is doing, you'll likely want someone else to be able to use it.
This may come in the form of a teammate or colleague trying to replicate and validate your results or through a customer using your model as part of a service or application you offer.
Saving a model also allows you to reuse it later without having to go through retraining it. Which is helpful, especially when your training times start to increase.
You can save a Scikit-Learn model using Python's in-built pickle module.
For larger models, it may be more efficient to use Joblib.
Woah!
We've covered a lot of ground fast...
Let's break things down a bit more by revisting each section.
1. Getting the data ready
Data doesn't always come ready to use with a Scikit-Learn machine learning model.
Three of the main steps you'll often have to take are:
Splitting the data into features (usually
X) and labels (usuallyy).Splitting the data into training and testing sets (and possibly a validation set).
Filling (also called imputing) or disregarding missing values.
Converting non-numerical values to numerical values (also call feature encoding).
Let's see an example.
Nice! Looks like our dataset has 303 samples with 13 features (13 columns).
Let's check out the labels.
Beautiful, 303 labels with values of 0 (no heart disease) and 1 (heart disease).
Now let's split our data into training and test sets, we'll use an 80/20 split (80% of samples for training and 20% of samples for testing).
1.1 Make sure it's all numerical
Computers love numbers.
So one thing you'll often have to make sure of is that your datasets are in numerical form.
This even goes for datasets which contain non-numerical features that you may want to include in a model.
For example, if we were working with a car sales dataset, how might we turn features such as Make and Colour into numbers?
Let's figure it out.
First, we'll import the car-sales-extended.csv dataset.
We can check the dataset types with .dtypes.
Notice the Make and Colour features are of dtype=object (they're strings) where as the rest of the columns are of dtype=int64.
If we want to use the Make and Colour features in our model, we'll need to figure out how to turn them into numerical form.
Now let's try and build a model on our car_sales data.
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
/var/folders/c4/qj4gdk190td18bqvjjh0p3p00000gn/T/ipykernel_23180/1044518071.py in ?()
1 # Try to predict with random forest on price column (doesn't work)
2 from sklearn.ensemble import RandomForestRegressor
3
4 model = RandomForestRegressor()
----> 5 model.fit(X_train, y_train)
6 model.score(X_test, y_test)
~/miniforge3/envs/ai/lib/python3.11/site-packages/sklearn/base.py in ?(estimator, *args, **kwargs)
1469 skip_parameter_validation=(
1470 prefer_skip_nested_validation or global_skip_validation
1471 )
1472 ):
-> 1473 return fit_method(estimator, *args, **kwargs)
~/miniforge3/envs/ai/lib/python3.11/site-packages/sklearn/ensemble/_forest.py in ?(self, X, y, sample_weight)
359 # Validate or convert input data
360 if issparse(y):
361 raise ValueError("sparse multilabel-indicator for y is not supported.")
362
--> 363 X, y = self._validate_data(
364 X,
365 y,
366 multi_output=True,
~/miniforge3/envs/ai/lib/python3.11/site-packages/sklearn/base.py in ?(self, X, y, reset, validate_separately, cast_to_ndarray, **check_params)
646 if "estimator" not in check_y_params:
647 check_y_params = {**default_check_params, **check_y_params}
648 y = check_array(y, input_name="y", **check_y_params)
649 else:
--> 650 X, y = check_X_y(X, y, **check_params)
651 out = X, y
652
653 if not no_val_X and check_params.get("ensure_2d", True):
~/miniforge3/envs/ai/lib/python3.11/site-packages/sklearn/utils/validation.py in ?(X, y, accept_sparse, accept_large_sparse, dtype, order, copy, force_writeable, force_all_finite, ensure_2d, allow_nd, multi_output, ensure_min_samples, ensure_min_features, y_numeric, estimator)
1297 raise ValueError(
1298 f"{estimator_name} requires y to be passed, but the target y is None"
1299 )
1300
-> 1301 X = check_array(
1302 X,
1303 accept_sparse=accept_sparse,
1304 accept_large_sparse=accept_large_sparse,
~/miniforge3/envs/ai/lib/python3.11/site-packages/sklearn/utils/validation.py in ?(array, accept_sparse, accept_large_sparse, dtype, order, copy, force_writeable, force_all_finite, ensure_2d, allow_nd, ensure_min_samples, ensure_min_features, estimator, input_name)
1009 )
1010 array = xp.astype(array, dtype, copy=False)
1011 else:
1012 array = _asarray_with_order(array, order=order, dtype=dtype, xp=xp)
-> 1013 except ComplexWarning as complex_warning:
1014 raise ValueError(
1015 "Complex data not supported\n{}\n".format(array)
1016 ) from complex_warning
~/miniforge3/envs/ai/lib/python3.11/site-packages/sklearn/utils/_array_api.py in ?(array, dtype, order, copy, xp, device)
747 # Use NumPy API to support order
748 if copy is True:
749 array = numpy.array(array, order=order, dtype=dtype)
750 else:
--> 751 array = numpy.asarray(array, order=order, dtype=dtype)
752
753 # At this point array is a NumPy ndarray. We convert it to an array
754 # container that is consistent with the input's namespace.
~/miniforge3/envs/ai/lib/python3.11/site-packages/pandas/core/generic.py in ?(self, dtype, copy)
2149 def __array__(
2150 self, dtype: npt.DTypeLike | None = None, copy: bool_t | None = None
2151 ) -> np.ndarray:
2152 values = self._values
-> 2153 arr = np.asarray(values, dtype=dtype)
2154 if (
2155 astype_is_view(values.dtype, arr.dtype)
2156 and using_copy_on_write()
ValueError: could not convert string to float: 'Honda'
Oh no! We get a another ValueError (some of data is in string format rather than numerical format).
Machine learning models prefer to work with numbers than text.
So we'll have to convert the non-numerical features into numbers first.
The process of turning categorical features into numbers is often referred to as encoding.
Scikit-Learn has a fantastic in-depth guide on Encoding categorical features.
But let's look at one of the most straightforward ways to turn categorical features into numbers, one-hot encoding.
In machine learning, one-hot encoding gives a value of 1 to the target value and a value of 0 to the other values.
For example, let's say we had five samples and three car make options, Honda, Toyota, BMW.
And our samples were:
Honda
BMW
BMW
Toyota
Toyota
If we were to one-hot encode these, it would look like:
| Sample | Honda | Toyota | BMW |
|---|---|---|---|
| 1 | 1 | 0 | 0 |
| 2 | 0 | 0 | 1 |
| 3 | 0 | 0 | 1 |
| 4 | 0 | 1 | 0 |
| 5 | 0 | 1 | 0 |
Notice how there's a 1 for each target value but a 0 for each other value.
We can use the following steps to one-hot encode our dataset:
Import
sklearn.preprocessing.OneHotEncoderto one-hot encode our features andsklearn.compose.ColumnTransformerto target the specific columns of our DataFrame to transform.Define the categorical features we'd like to transform.
Create an instance of the
OneHotEncoder.Create an instance of
ColumnTransformerand feed it the transforms we'd like to make.Fit the instance of the
ColumnTransformerto our data and transform it with thefit_transform(X)method.
Note: In Scikit-Learn, the term "transformer" is often used to refer to something that transforms data.
Note: You might be thinking why we considered
Doorsas a categorical variable. Which is a good question consideringDoorsis already numerical. Well, the answer is thatDoorscould be either numerical or categorical. However, I've decided to go with categorical, since where I'm from, number of doors is often a different category of car. For example, you can shop for 4-door cars or shop for 5-door cars (which always confused me since where's the 5th door?). However, you could experiment with treating this value as numerical or categorical, training a model on each, and then see how each model performs.
Woah! Looks like our samples are all numerical, what did our data look like previously?
It seems OneHotEncoder and ColumnTransformer have turned all of our data samples into numbers.
Let's check out the first transformed sample.
And what were these values originally?
1.1.1 Nuemrically encoding data with pandas
Another way we can numerically encode data is directly with pandas.
We can use the pandas.get_dummies() (or pd.get_dummies() for short) method and then pass it our target columns.
In return, we'll get a one-hot encoded version of our target columns.
Let's remind ourselves of what our DataFrame looks like.
Wonderful, now let's use pd.get_dummies() to turn our categorical variables into one-hot encoded variables.
Nice!
Notice how there's a new column for each categorical option (e.g. Make_BMW, Make_Honda, etc).
But also notice how it also missed the Doors column?
This is because Doors is already numeric, so for pd.get_dummies() to work on it, we can change it to type object.
By default, pd.get_dummies() also turns all of the values to bools (True or False).
We can get the returned values as 0 or 1 by setting dtype=float.
Woohoo!
We've now turned our data into fully numeric form using Scikit-Learn and pandas.
Now you might be wondering...
Should you use Scikit-Learn or pandas for turning data into numerical form?
And the answer is either.
But as a rule of thumb:
If you're performing quick data analysis and running small modelling experiments, use
pandasas it's generally quite fast to get up and running.If you're performing a larger scale modelling experiment or would like to put your data processing steps into a production pipeline, I'd recommend leaning towards Scikit-Learn, specifically a Scikit-Learn Pipeline (chaining together multiple estimator/modelling steps).
Since we've turned our data into numerical form, how about we try and fit our model again?
Let's recreate a train/test split except this time we'll use transformed_X instead of X.
1.2 What if there were missing values in the data?
Holes in the data means holes in the patterns your machine learning model can learn.
Many machine learning models don't work well or produce errors when they're used on datasets with missing values.
A missing value can appear as a blank, as a NaN or something similar.
There are two main options when dealing with missing values:
Fill them with some given or calculated value (imputation) - For example, you might fill missing values of a numerical column with the mean of all the other values. The practice of calculating or figuring out how to fill missing values in a dataset is called imputing. For a great resource on imputing missing values, I'd recommend refering to the Scikit-Learn user guide.
Remove them - If a row or sample has missing values, you may opt to remove them from your dataset completely. However, this potentially results in using less data to build your model.
Note: Dealing with missing values differs from problem to problem, meaning there's no 100% best way to fill missing values across datasets and problem types. It will often take careful experimentation and practice to figure out the best way to deal with missing values in your own datasets.
To practice dealing with missing values, let's import a version of the car_sales dataset with several missing values (namely car-sales-extended-missing-data.csv).
Notice the NaN value in row 7 for the Odometer (KM) column, that means pandas has detected a missing value there.
However, if you're dataset is large, it's likely you aren't going to go through it sample by sample to find the missing values.
Luckily, pandas has a method called pd.DataFrame.isna() which is able to detect missing values.
Let's try it on our DataFrame.
Hmm... seems there's about 50 or so missing values per column.
How about we try and split the data into features and labels, then convert the categorical data to numbers, then split the data into training and test and then try and fit a model on it (just like we did before)?
Now we can convert the categorical columns into one-hot encodings (just as before).
Finally, let's split the missing data samples into train and test sets and then try to fit and score a model on them.
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
Cell In[52], line 8
6 # Fit and score a model
7 model = RandomForestRegressor()
----> 8 model.fit(X_train, y_train)
9 model.score(X_test, y_test)
, in _fit_context.<locals>.decorator.<locals>.wrapper(estimator, *args, **kwargs)
1466 estimator._validate_params()
1468 with config_context(
1469 skip_parameter_validation=(
1470 prefer_skip_nested_validation or global_skip_validation
1471 )
1472 ):
-> 1473 return fit_method(estimator, *args, **kwargs)
, in BaseForest.fit(self, X, y, sample_weight)
360 if issparse(y):
361 raise ValueError("sparse multilabel-indicator for y is not supported.")
--> 363 X, y = self._validate_data(
364 X,
365 y,
366 multi_output=True,
367 accept_sparse="csc",
368 dtype=DTYPE,
369 force_all_finite=False,
370 )
371 # _compute_missing_values_in_feature_mask checks if X has missing values and
372 # will raise an error if the underlying tree base estimator can't handle missing
373 # values. Only the criterion is required to determine if the tree supports
374 # missing values.
375 estimator = type(self.estimator)(criterion=self.criterion)
, in BaseEstimator._validate_data(self, X, y, reset, validate_separately, cast_to_ndarray, **check_params)
648 y = check_array(y, input_name="y", **check_y_params)
649 else:
--> 650 X, y = check_X_y(X, y, **check_params)
651 out = X, y
653 if not no_val_X and check_params.get("ensure_2d", True):
, in check_X_y(X, y, accept_sparse, accept_large_sparse, dtype, order, copy, force_writeable, force_all_finite, ensure_2d, allow_nd, multi_output, ensure_min_samples, ensure_min_features, y_numeric, estimator)
1297 raise ValueError(
1298 f"{estimator_name} requires y to be passed, but the target y is None"
1299 )
1301 X = check_array(
1302 X,
1303 accept_sparse=accept_sparse,
(...)
1315 input_name="X",
1316 )
-> 1318 y = _check_y(y, multi_output=multi_output, y_numeric=y_numeric, estimator=estimator)
1320 check_consistent_length(X, y)
1322 return X, y
, in _check_y(y, multi_output, y_numeric, estimator)
1326 """Isolated part of check_X_y dedicated to y validation"""
1327 if multi_output:
-> 1328 y = check_array(
1329 y,
1330 accept_sparse="csr",
1331 force_all_finite=True,
1332 ensure_2d=False,
1333 dtype=None,
1334 input_name="y",
1335 estimator=estimator,
1336 )
1337 else:
1338 estimator_name = _check_estimator_name(estimator)
, in check_array(array, accept_sparse, accept_large_sparse, dtype, order, copy, force_writeable, force_all_finite, ensure_2d, allow_nd, ensure_min_samples, ensure_min_features, estimator, input_name)
1058 raise ValueError(
1059 "Found array with dim %d. %s expected <= 2."
1060 % (array.ndim, estimator_name)
1061 )
1063 if force_all_finite:
-> 1064 _assert_all_finite(
1065 array,
1066 input_name=input_name,
1067 estimator_name=estimator_name,
1068 allow_nan=force_all_finite == "allow-nan",
1069 )
1071 if copy:
1072 if _is_numpy_namespace(xp):
1073 # only make a copy if `array` and `array_orig` may share memory`
, in _assert_all_finite(X, allow_nan, msg_dtype, estimator_name, input_name)
120 if first_pass_isfinite:
121 return
--> 123 _assert_all_finite_element_wise(
124 X,
125 xp=xp,
126 allow_nan=allow_nan,
127 msg_dtype=msg_dtype,
128 estimator_name=estimator_name,
129 input_name=input_name,
130 )
, in _assert_all_finite_element_wise(X, xp, allow_nan, msg_dtype, estimator_name, input_name)
155 if estimator_name and input_name == "X" and has_nan_error:
156 # Improve the error message on how to handle missing values in
157 # scikit-learn.
158 msg_err += (
159 f"\n{estimator_name} does not accept missing values"
160 " encoded as NaN natively. For supervised learning, you might want"
(...)
170 "#estimators-that-handle-nan-values"
171 )
--> 172 raise ValueError(msg_err)
ValueError: Input y contains NaN.
Ahh... dam! Another ValueError (our input data contains missing values).
ValueError: Input y contains NaN.
Looks like the model we're trying to use doesn't work with missing values.
When we try to fit it on a dataset with missing samples, Scikit-Learn produces an error similar to:
ValueError: Input X contains NaN. RandomForestRegressor does not accept missing values encoded as NaN natively...
Looks like if we want to use RandomForestRegressor, we'll have to either fill or remove the missing values.
Note: Scikit-Learn does have a list of models which can handle NaNs or missing values directly.
Such as,
sklearn.ensemble.HistGradientBoostingClassifier
or
sklearn.ensemble.HistGradientBoostingRegressor.
As an experiment, you may want to try the following:
from sklearn.ensemble import HistGradientBoostingRegressor# Try a model that can handle NaNs natively nan_model = HistGradientBoostingRegressor() nan_model.fit(X_train, y_train) nan_model.score(X_test, y_test)
How can fill (impute) or remove these?
1.2.1 Fill missing data with pandas
Let's see how we might fill missing values with pandas.
For categorical values, one of the simplest ways is to fill the missing fields with the string "missing".
We could do this for the Make and Colour features.
As for the Doors feature, we could use "missing" or we could fill it with the most common option of 4.
With the Odometer (KM) feature, we can use the mean value of all the other values in the column.
And finally, for those samples which are missing a Price value, we can remove them (since Price is the target value, removing probably causes less harm than imputing, however, you could design an experiment to test this).
In summary:
| Column/Feature | Fill missing value with | | ----- | ----- | | Make | "missing" | | Colour | "missing" | | Doors | 4 (most common value) | | Odometer (KM) | mean of Odometer (KM) | | Price (target) | NA, remove samples missing Price |
Note: The practice of filling missing data with given or calculated values is called imputation. And it's important to remember there's no perfect way to fill missing data (unless it's with data that should've actually been there in the first place). The methods we're using are only one of many. The techniques you use will depend heavily on your dataset. A good place to look would be searching for "data imputation techniques".
Let's start with the Make column.
We can use the pandas method fillna(value="missing", inplace=True) to fill all the missing values with the string "missing".
And we can do the same with the Colour column.
How many missing values do we have now?
Next, we'll fill the Odometer (KM) column with the mean value of itself.
How many missing values do we have now?
Woohoo! That's looking a lot better.
Finally, we can remove the rows which are missing the target value Price.
Note: Another option would be to impute the
Pricevalue with the mean or median or some other calculated value (such as by using similar cars to estimate the price), however, to keep things simple and prevent introducing too many fake labels to the data, we'll remove the samples missing aPricevalue.
We can remove rows with missing values in place from a pandas DataFrame with the pandas.DataFrame.dropna(inplace=True) method.
That should be no more missing values!
Since we removed samples missing a Price value, there's now less overall samples but none of them have missing values.
Can we fit a model now?
Let's try!
First we'll create the features and labels.
Then we'll convert categorical variables into numbers via one-hot encoding.
Then we'll split the data into training and test sets just like before.
Finally, we'll try to fit a RandomForestRegressor() model to the newly filled data.
Fantastic!!!
Looks like filling the missing values with pandas worked!
Our model can be fit to the data without issues.
1.2.2 Filling missing data and transforming categorical data with Scikit-Learn
Now we've filled the missing columns using pandas functions, you might be thinking, "Why pandas? I thought this was a Scikit-Learn introduction?".
Not to worry, Scikit-Learn provides a class called sklearn.impute.SimpleImputer() which allows us to do a similar thing.
SimpleImputer() transforms data by filling missing values with a given strategy parameter.
And we can use it to fill the missing values in our DataFrame as above.
At the moment, our dataframe has no mising values.
Let's reimport it so it has missing values and we can fill them with Scikit-Learn.
To begin, we'll remove the rows which are missing a Price value.
Now there are no rows missing a Price value.
Since we don't have to fill any Price values, let's split our data into features (X) and labels (y).
We'll also split the data into training and test sets.
Note: We've split the data into train & test sets here first to perform filling missing values on them separately. This is best practice as the test set is supposed to emulate data the model has never seen before. For categorical variables, it's generally okay to fill values across the whole dataset. However, for numerical vairables, you should only fill values on the test set that have been computed from the training set.
Training and test sets created!
Let's now setup a few instances of SimpleImputer() to fill various missing values.
We'll use the following strategies and fill values:
For categorical columns (
Make,Colour),strategy="constant",fill_value="missing"(fill the missing samples with a consistent value of"missing".For the
Doorcolumn,strategy="constant",fill_value=4(fill the missing samples with a consistent value of4).For the numerical column (
Odometer (KM)),strategy="mean"(fill the missing samples with the mean of the target column).Note: There are more
strategyand fill options in theSimpleImputer()documentation.
Imputers created!
Now let's define which columns we'd like to impute on.
Why?
Because we'll need these shortly (I'll explain in the next text cell).
Columns defined!
Now how might we transform our columns?
Hint: we can use the sklearn.compose.ColumnTransformer class from Scikit-Learn, in a similar way to what we did before to get our data to all numeric values.
That's the reason we defined the columns we'd like to transform.
So we can use the ColumnTransformer() class.
ColumnTransformer() takes as input a list of tuples in the form (name_of_transform, transformer_to_use, columns_to_transform) specifying which columns to transform and how to transform them.
For example:
In this case, the variables in the tuple are:
name_of_transform="cat_imputer"transformer_to_use=cat_imputer(the instance ofSimpleImputer()we defined above)columns_to_transform=categorical_features(the list of categorical features we defined above).
Let's exapnd upon this by extending the example.
Nice!
The next step is to fit our ColumnTransformer() instance (imputer) to the training data and transform the testing data.
In other words we want to:
Learn the imputation values from the training set.
Fill the missing values in the training set with the values learned in 1.
Fill the missing values in the testing set with the values learned in 1.
Why this way?
In our case, we're not calculating many variables (except the mean of the Odometer (KM) column), however, remember that the test set should always remain as unseen data.
So when filling values in the test set, they should only be with values calculated or imputed from the training sets.
We can achieve steps 1 & 2 simultaneously with the ColumnTransformer.fit_transform() method (fit = find the values to fill, transform = fill them).
And then we can perform step 3 with the ColumnTransformer.transform() method (we only want to transform the test set, not learn different values to fill).
Wonderful!
Let's now turn our filled_X_train and filled_X_test arrays into DataFrames to inspect their missing values.
And is there any missing data in the test set?
What about the original?
Perfect!
No more missing values!
But wait...
Is our data all numerical?
Ahh... looks like our Make and Colour columns are still strings.
Let's one-hot encode them along with the Doors column to make sure they're numerical, just as we did previously.
Nice!
Now our data is:
All numerical
No missing values
Let's try and fit a model!
You might have noticed this result is slightly different to before.
Why do you think this is?
It's because we've created our training and testing sets differently.
We split the data into training and test sets before filling the missing values.
Previously, we did the reverse, filled missing values before splitting the data into training and test sets.
Doing this can lead to information from the training set leaking into the testing set.
Remember, one of the most important concepts in machine learning is making sure your model doesn't see any testing data before evaluation.
We'll keep practicing but for now, some of the main takeaways are:
Keep your training and test sets separate.
Most datasets you come across won't be in a form ready to immediately start using them with machine learning models. And some may take more preparation than others to get ready to use.
For most machine learning models, your data has to be numerical. This will involve converting whatever you're working with into numbers. This process is often referred to as feature engineering or feature encoding.
Some machine learning models aren't compatible with missing data. The process of filling missing data is referred to as data imputation.
2. Choosing the right estimator/algorithm for your problem
Once you've got your data ready, the next step is to choose an appropriate machine learning algorithm or model to find patterns in your data.
Some things to note:
Scikit-Learn refers to machine learning models and algorithms as estimators.
Classification problem - predicting a category (heart disease or not).
Sometimes you'll see
clf(short for classifier) used as a classification estimator instance's variable name.
Regression problem - predicting a number (selling price of a car).
Unsupervised problem (data with no labels) - clustering (grouping unlabelled samples with other similar unlabelled samples).
If you know what kind of problem you're working with, one of the next places you should look at is the Scikit-Learn algorithm cheatsheet.
This cheatsheet gives you a bit of an insight into the algorithm you might want to use for the problem you're working on.
It's important to remember, you don't have to explicitly know what each algorithm is doing on the inside to start using them.
If you start to apply different algorithms but they don't seem to be working (not performing as well as you'd like), that's when you'd start to look deeper into each one.
Let's check out the cheatsheet and follow it for some of the problems we're working on.
You can see it's split into four main categories. Regression, classification, clustering and dimensionality reduction. Each has their own different purpose but the Scikit-Learn team has designed the library so the workflows for each are relatively similar.
2.1 Picking a machine learning model for a regression problem
Let's start with a regression problem (trying to predict a number). We'll use the California Housing dataset built into Scikit-Learn's datasets module.
The goal of the California Housing dataset is to predict a given district's median house value (in hundreds of thousands of dollars) on things like the age of the home, the number of rooms, the number of bedrooms, number of people living the home and more.
Since it's in a dictionary, let's turn it into a DataFrame so we can inspect it better.
Beautiful, our goal here is to use the feature columns, such as:
MedInc- median income in block groupHouseAge- median house age in block groupAveRooms- average number of rooms per householdAveBedrms- average number of bedrooms per household
To predict the target column which expresses the median house value for specfici California districts in hundreds of thousands of dollars (e.g. 4.526 = $452,600).
In essence, each row is a different district in California (the data) and we're trying to build a model to predict the median house value in that distract (the target/label) given a series of attributes about the houses in that district.
Since we have data and labels, this is a supervised learning problem.
And since we're trying to predict a number, it's a regression problem.
Knowing these two things, how do they line up on the Scikit-Learn machine learning algorithm cheat-sheet?
Following the map through, knowing what we know, it suggests we try RidgeRegression. Let's chek it out.
What if RidgeRegression didn't work? Or what if we wanted to improve our results?
Following the diagram, the next step would be to try EnsembleRegressors.
Ensemble is another word for multiple models put together to make a decision.
One of the most common and useful ensemble methods is the Random Forest. Known for its fast training and prediction times and adaptibility to different problems.
The basic premise of the Random Forest is to combine a number of different decision trees, each one random from the other and make a prediction on a sample by averaging the result of each decision tree.
An in-depth discussion of the Random Forest algorithm is beyond the scope of this notebook but if you're interested in learning more, An Implementation and Explanation of the Random Forest in Python by Will Koehrsen is a great read.
Since we're working with regression, we'll use Scikit-Learn's RandomForestRegressor.
We can use the exact same workflow as above. Except for changing the model.
Woah!
We get a good boost in score on the test set by changing the model.
This is another incredibly important concept in machine learning, if at first something doesn't achieve what you'd like, experiment, experiment, experiment!
At first, the Scikit-Learn algorithm diagram can seem confusing.
But once you get a little practice applying different models to different problems, you'll start to pick up which sorts of algorithms do better with different types of data.
2.2 Picking a machine learning model for a classification problem
Now, let's check out the choosing process for a classification problem.
Say you were trying to predict whether or not a patient had heart disease based on their medical records.
The dataset in ../data/heart-disease.csv (or at heart-disease.csv) contains data for just that problem.
Similar to the California Housing dataset, here we want to use all of the available data to predict the target column (1 for if a patient has heart disease and 0 for if they don't).
So what do we know?
We've got 303 samples (1 row = 1 sample) and we're trying to predict whether or not a patient has heart disease.
Because we're trying to predict whether each sample is one thing or another, we've got a classification problem.
Let's see how it lines up with our Scikit-Learn algorithm cheat-sheet.
Following the cheat-sheet we end up at LinearSVC which stands for Linear Support Vector Classifier. Let's try it on our data.
Straight out of the box (with no tuning or improvements) our model achieves over 85% accruacy!
Although this is a sensational result to begin with, let's check out the diagram and see what other models we might use.
Following the path (and skipping a few, don't worry, we'll get to this) we come up to EnsembleMethods again.
Except this time, we'll be looking at ensemble classifiers instead of regressors.
Remember our RandomForestRegressor from above?
We'll it has a dance partner, RandomForestClassifier which is an ensemble based machine model learning model for classification.
You might be able to guess what we can use it for (hint: classification problems).
Let's try!
Hmmm, it looks like the default hyperparameters of RandomForestClassifier don't perform as well as LinearSVC.
Other than trying another classification model, we could start to run experiments to try and improve these models via hyperparameter tuning.
Hyperparameter tuning is fancy term for adjusting some settings on a model to try and make it better.
It usually happens once you've found a decent baseline model that you'd like to improve upon.
In this case, we could take either the RandomForestClassifier or the LinearSVC and try and improve it with hyperparameter tuning (which we'll see later on).
For example, you could try and take the n_estimators parameter (the number of trees in the forest) of RandomForestClassifier and change it from 100 (default) to 1000 and see what happens.
What about the other models?
Looking at the Scikit-Learn aglorithm cheat-sheet and the examples above, you may have noticed we've skipped a few models.
Why?
The first reason is time.
Covering every single one would take a fair bit longer than what we've done here. And the second one is the effectiveness of ensemble methods.
A little tidbit for modelling in machine learning is:
If you have structured data (tables, spreadsheets or dataframes), use ensemble methods, such as, a Random Forest.
If you have unstructured data (text, images, audio, things not in tables), use deep learning or transfer learning (see the ZTM TensorFlow and PyTorch courses for more on deep learning).
For this notebook, we're focused on structured data, which is why the Random Forest has been our model of choice.
If you'd like to learn more about the Random Forest and why it's the war horse of machine learning, check out these resources:
Experiment until something works
The beautiful thing is, the way the Scikit-Learn API is designed, once you know the way with one model, using another is much the same.
And since a big part of being a machine learning engineer or data scientist is experimenting, you might want to try out some of the other models on the cheat-sheet and see how you go. The more you can reduce the time between experiments, the better.
3. Fit the model to data and using it to make predictions
Now you've chosen a model, the next step is to have it learn from the data so it can be used for predictions in the future.
If you've followed through, you've seen a few examples of this already.
3.1 Fitting a model to data
In Scikit-Learn, the process of having a machine learning model learn patterns from a dataset involves calling the fit() method and passing it data, such as, fit(X, y).
Where X is a feature array and y is a target array.
Other names for X include:
Data
Feature variables
Features
Other names for y include:
Labels
Target variable
For supervised learning there is usually an X and y.
For unsupervised learning, there's no y (no labels).
Let's revisit the example of using patient data (X) to predict whether or not they have heart disease (y).
What's happening here?
Calling the fit() method will cause the machine learning algorithm to attempt to find patterns between X and y. Or if there's no y, it'll only find the patterns within X.
Let's see X.
And y.
Passing X and y to fit() will cause the model to go through all of the examples in X (data) and see what their corresponding y (label) is.
How the model does this is different depending on the model you use.
Explaining the details of each would take an entire textbook.
For now, you could imagine it similar to how you would figure out patterns if you had enough time.
You'd look at the feature variables, X, the age, sex, chol (cholesterol) and see what different values led to the labels, y, 1 for heart disease, 0 for not heart disease.
This concept, regardless of the problem, is similar throughout all of machine learning.
During training (finding patterns in data):
A machine learning algorithm looks at a dataset, finds patterns, tries to use those patterns to predict something and corrects itself as best it can with the available data and labels. It stores these patterns for later use.
During testing or in production (using learned patterns):
A machine learning algorithm uses the patterns its previously learned in a dataset to make a prediction on some unseen data.
3.2 Making predictions using a machine learning model
Now we've got a trained model, one which has hoepfully learned patterns in the data, you'll want to use it to make predictions.
Scikit-Learn enables this in several ways.
Two of the most common and useful are predict() and predict_proba().
Let's see them in action.
Given data in the form of X, the predict() function returns labels in the form of y.
Note: For the
predict()function to work, it must be passedX(data) in the same format the model was trained on. For example, if a model was trained on 10 features formatted in a certain way, predictions should be made on data with 10 features fortmatted in a certain way. Anything different and it will return an error.
It's standard practice to save these predictions to a variable named something like y_preds for later comparison to y_test or y_true (usually same as y_test just another name).
Another way evaluating predictions (comparing them to the truth labels) is with Scikit-Learn's sklearn.metrics module.
Inside, you'll find method such as accuracy_score(), which is the default evaluation metric for classification problems.
predict_proba() returns the probabilities (proba is short for probability) of a classification label.
Let's see the difference.
predict_proba() returns an array of five arrays each containing two values.
Each number is the probability of a label given a sample.
This output means for the sample X_test[:1], the model is predicting label 0 (index 0) with a probability score of 0.9.
Because the highest probability score is at index 0 (and it's over 0.5), when using predict(), a label of 0 is assigned.
Where does 0.5 come from?
Because our problem is a binary classification task (heart disease or not heart disease), predicting a label with 0.5 probability every time would be the same as a coin toss (guessing 50/50 every time).
Therefore, once the prediction probability of a sample passes 0.5 for a certain label, it's assigned that label.
predict() can also be used for regression models.
Now we can evaluate our regression model by using sklearn.metrics.mean_absolute_error which returns the average error across all samples.
Now we've seen how to get a model how to find patterns in data using the fit() function and make predictions using what its learned using the predict() and predict_proba() functions, it's time to evaluate those predictions.
4. Evaluating a model
Once you've trained a model, you'll want a way to measure how trustworthy its predictions are.
Across the board, the main idea of evaluating a model is to compare the model's predictions to what they should've ideally been (the truth labels).
Scikit-Learn implements 3 different methods of evaluating models.
The
score()method. Callingscore()on a model instance will return a metric assosciated with the type of model you're using. The metric depends on which model you're using.The
scoringparameter. This parameter can be passed to methods such ascross_val_score()orGridSearchCV()to tell Scikit-Learn to use a specific type of scoring metric.Problem-specific metric functions available in
sklearn.metrics. Similar to how thescoringparameter can be passed different scoring functions, Scikit-Learn implements these as stand alone functions.
The scoring function you use will also depend on the problem you're working on.
Classification problems have different evaluation metrics and scoring functions to regression problems.
Let's look at some examples.
4.1 General model evaluation with score()
If we bring down the code from our previous classification problem (building a classifier to predict whether or not someone has heart disease based on their medical records).
We can see the score() method come into play.
Once the model has been fit on the training data (X_train, y_train), we can call the score() method on it and evaluate our model on the test data, data the model has never seen before (X_test, y_test).
Each model in Scikit-Learn implements a default metric for score() which is suitable for the problem.
For example:
Classifier models generally use
metrics.accuracy_score()as the defaultscore()metric.Regression models generally use
metrics.r2_scoreas the defaultscore()metric.There many more classification and regression specific metrics implemented in
sklearn.metrics.
Because clf is an instance of RandomForestClassifier, the score() method uses mean accuracy as its score method.
You can find this by pressing SHIFT + TAB (inside a Jupyter Notebook, may be different elsewhere) within the brackets of score() when called on a model instance.
Behind the scenes, score() makes predictions on X_test using the trained model and then compares those predictions to the actual labels y_test.
A classification model which predicts everything 100% correct would receive an accuracy score of 1.0 (or 100%).
Our model doesn't get everything correct, but at ~85% accuracy (0.85 * 100), it's still far better than guessing.
Let's do the same but with the regression code from above.
Due to the consistent design of the Scikit-Learn library, we can call the same score() method on model.
Here, model is an instance of RandomForestRegressor.
And since it's a regression model, the default metric built into score() is the coefficient of determination or R^2 (pronounced R-sqaured).
Remember, you can find this by pressing SHIFT + TAB within the brackets of score() when called on a model instance.
The best possible value here is 1.0, this means the model predicts the target regression values exactly.
Calling the score() method on any model instance and passing it test data is a good quick way to see how your model is going.
However, when you get further into a problem, it's likely you'll want to start using more powerful metrics to evaluate your models performance.
4.2 Evaluating your models using the scoring parameter
The next step up from using score() is to use a custom scoring parameter with cross_val_score() or GridSearchCV.
As you may have guessed, the scoring parameter you set will be different depending on the problem you're working on.
We'll see some specific examples of different parameters in a moment but first let's check out cross_val_score().
To do so, we'll copy the heart disease classification code from above and then add another line at the top.
Using cross_val_score() is slightly different to score().
Let's see a code example first and then we'll go through the details.
What's happening here?
The first difference you might notice is cross_val_score() returns an array where as score() only returns a single number.
cross_val_score() returns an array because of a parameter called cv, which stands for cross-validation.
When cv isn't set, cross_val_score() will return an array of 5 numbers by default (cv=None is the same as setting cv=5).
Remember, you can see the parameters of a function using SHIFT + TAB (inside a Jupyter Notebook) from within the brackets.
But wait, you might be thinking, what even is cross-validation?
A visual might be able to help.
We've dealt with Figure 1.0 before using score(X_test, y_test).
But looking deeper into this, if a model is trained using the training data or 80% of samples, this means 20% of samples aren't used for the model to learn anything.
This also means depending on what 80% is used to train on and what 20% is used to evaluate the model, it may achieve a score which doesn't reflect the entire dataset.
For example, if a lot of easy examples are in the 80% training data, when it comes to test on the 20%, your model may perform poorly.
The same goes for the reverse.
Figure 2.0 shows 5-fold cross-validation, a method which tries to provide a solution to:
Not training on all the data (always keeping training and test sets separate).
Avoiding getting lucky scores on single splits of the data.
Instead of training only on 1 training split and evaluating on 1 testing split, 5-fold cross-validation does it 5 times.
On a different split each time, returning a score for each.
Why 5-fold?
The actual name of this setup K-fold cross-validation. Where K is an abitrary number. We've used 5 because it looks nice visually, and it is the default value in sklearn.model_selection.cross_val_score.
Figure 2.0 is what happens when we run the following.
Since we set cv=5 (5-fold cross-validation), we get back 5 different scores instead of 1.
Taking the mean of this array gives us a more in-depth idea of how our model is performing by converting the 5 scores into one.
Notice, the average cross_val_score() is slightly lower than single value returned by score().
In this case, if you were asked to report the accuracy of your model, even though it's lower, you'd prefer the cross-validated metric over the non-cross-validated metric.
Wait?
We haven't used the scoring parameter at all.
By default, it's set to None.
Note: If you notice different scores each time you call
cross_val_score, this is because each data split is random every time. So the model may achieve higher/lower scores on different splits of the data. To get reproducible scores, you can set the random seed.
When scoring is set to None (by default), it uses the same metric as score() for whatever model is passed to cross_val_score().
In this case, our model is clf which is an instance of RandomForestClassifier which uses mean accuracy as the default score() metric.
You can change the evaluation score cross_val_score() uses by changing the scoring parameter.
And as you might have guessed, different problems call for different evaluation scores.
The Scikit-Learn documentation outlines a vast range of evaluation metrics for different problems but let's have a look at a few.
4.2.1 Classification model evaluation metrics
Four of the main evaluation metrics/methods you'll come across for classification models are:
Area under ROC curve (receiver operating characteristic curve)
Let's have a look at each of these. We'll bring down the classification code from above to go through some examples.
Accuracy
Accuracy is the default metric for the score() function within each of Scikit-Learn's classifier models. And it's probably the metric you'll see most often used for classification problems.
However, we'll see in a second how it may not always be the best metric to use.
Scikit-Learn returns accuracy as a decimal but you can easily convert it to a percentage.
Area Under Receiver Operating Characteristic (ROC) Curve
If this one sounds like a mouthful, its because reading the full name is.
It's usually referred to as AUC for Area Under Curve and the curve they're talking about is the Receiver Operating Characteristic or ROC for short.
So if hear someone talking about AUC or ROC, they're probably talking about what follows.
ROC curves are a comparison of true postive rate (tpr) versus false positive rate (fpr).
For clarity:
True positive = model predicts 1 when truth is 1
False positive = model predicts 1 when truth is 0
True negative = model predicts 0 when truth is 0
False negative = model predicts 0 when truth is 1
Now we know this, let's see one. Scikit-Learn lets you calculate the information required for a ROC curve using the roc_curve function.
Looking at these on their own doesn't make much sense. It's much easier to see their value visually.
Let's create a helper function to make a ROC curve given the false positive rates (fpr) and true positive rates (tpr).
Note: As of Scikit-Learn 1.2+, there is functionality of plotting a ROC curve. You can find this under
sklearn.metrics.RocCurveDisplay.
Looking at the plot for the first time, it might seem a bit confusing.
The main thing to take away here is our model is doing far better than guessing.
A metric you can use to quantify the ROC curve in a single number is AUC (Area Under Curve).
Scikit-Learn implements a function to caculate this called sklearn.metrics.roc_auc_score.
The maximum ROC AUC score you can achieve is 1.0 and generally, the closer to 1.0, the better the model.
I'll let you in a secret...
Although it was good practice, we didn't actually need to create our own plot_roc_curve function.
Scikit-Learn allows us to plot a ROC curve directly from our estimator/model by using the class method sklearn.metrics.RocCurveDisplay.from_estimator and passing it our estimator, X_test and y_test.
The most ideal position for a ROC curve to run along the top left corner of the plot.
This would mean the model predicts only true positives and no false positives. And would result in a ROC AUC score of 1.0.
You can see this by creating a ROC curve using only the y_test labels.
In reality, a perfect ROC curve is unlikely.
Confusion matrix
Another fantastic way to evaluate a classification model is by using a confusion matrix.
A confusion matrix is a quick way to compare the labels a model predicts and the actual labels it was supposed to predict.
In essence, giving you an idea of where the model is getting confused.
Again, this is probably easier visualized.
One way to do it is with pd.crosstab().
Creating a confusion matrix using Scikit-Learn
Scikit-Learn has multiple different implementations of plotting confusion matrices:
sklearn.metrics.ConfusionMatrixDisplay.from_estimator(estimator, X, y)- this takes a fitted estimator (like ourclfmodel), features (X) and labels (y), it then uses the trained estimator to make predictions onXand compares the predictions toyby displaying a confusion matrix.sklearn.metrics.ConfusionMatrixDisplay.from_predictions(y_true, y_pred)- this takes truth labels and predicted labels and compares them by displaying a confusion matrix.
Note: Both of these methods/classes require Scikit-Learn 1.0+. To check your version of Scikit-Learn run:
If you don't have 1.0+, you can upgrade at: https://scikit-learn.org/stable/install.html
Classification report
The final major metric you should consider when evaluating a classification model is a classification report.
A classification report is more so a collection of metrics rather than a single one.
You can create a classification report using Scikit-Learn's sklearn.metrics.classification_report` method.
Let's see one.
It returns four columns: precision, recall, f1-score and support.
The number of rows will depend on how many different classes there are. But there will always be three rows labell accuracy, macro avg and weighted avg.
Each term measures something slightly different:
Precision - Indicates the proportion of positive identifications (model predicted class
1) which were actually correct. A model which produces no false positives has a precision of 1.0.Recall - Indicates the proportion of actual positives which were correctly classified. A model which produces no false negatives has a recall of 1.0.
F1 score - A combination of precision and recall. A perfect model achieves an F1 score of 1.0.
Support - The number of samples each metric was calculated on.
Accuracy - The accuracy of the model in decimal form. Perfect accuracy is equal to 1.0, in other words, getting the prediction right 100% of the time.
Macro avg - Short for macro average, the average precision, recall and F1 score between classes. Macro avg doesn't take class imbalance into effect. So if you do have class imbalances (more examples of one class than another), you should pay attention to this.
Weighted avg - Short for weighted average, the weighted average precision, recall and F1 score between classes. Weighted means each metric is calculated with respect to how many samples there are in each class. This metric will favour the majority class (e.g. it will give a high value when one class out performs another due to having more samples).
When should you use each?
It can be tempting to base your classification models perfomance only on accuracy. And accuracy is a good metric to report, except when you have very imbalanced classes.
For example, let's say there were 10,000 people. And 1 of them had a disease. You're asked to build a model to predict who has it.
You build the model and find your model to be 99.99% accurate. Which sounds great! ...until you realise, all its doing is predicting no one has the disease, in other words all 10,000 predictions are false.
In this case, you'd want to turn to metrics such as precision, recall and F1 score.
You can see here, we've got an accuracy of 0.9999 (99.99%), great precision and recall on class 0.0 but nothing for class 1.0.
Ask yourself, although the model achieves 99.99% accuracy, is it useful?
To summarize:
Accuracy is a good measure to start with if all classes are balanced (e.g. same amount of samples which are labelled with 0 or 1)
Precision and recall become more important when classes are imbalanced.
If false positive predictions are worse than false negatives, aim for higher precision.
If false negative predictions are worse than false positives, aim for higher recall.
Resource: For more on precision and recall and the tradeoffs between them, I'd suggest going through the Scikit-Learn Precision-Recall guide.
4.2.2 Regression model evaluation metrics
Similar to classification, there are several metrics you can use to evaluate your regression models.
We'll check out the following.
R^2 (pronounced r-squared) or coefficient of determination - Compares your models predictions to the mean of the targets. Values can range from negative infinity (a very poor model) to 1. For example, if all your model does is predict the mean of the targets, its R^2 value would be 0. And if your model perfectly predicts a range of numbers it's R^2 value would be 1. Higher is better.
Mean absolute error (MAE) - The average of the absolute differences between predictions and actual values. It gives you an idea of how wrong your predictions were. Lower is better.
Mean squared error (MSE) - The average squared differences between predictions and actual values. Squaring the errors removes negative errors. It also amplifies outliers (samples which have larger errors). Lower is better.
Let's see them in action. First, we'll bring down our regression model code again.
R^2 Score (coefficient of determination)
Once you've got a trained regression model, the default evaluation metric in the score() function is R^2.
Outside of the score() function, R^2 can be calculated using Scikit-Learn's r2_score() function.
A model which only predicted the mean would get a score of 0.
And a perfect model would get a score of 1.
For your regression models, you'll want to maximise R^2, whilst minimising MAE and MSE.
Mean Absolute Error (MAE)
A model's mean absolute error can be calculated with Scikit-Learn's sklearn.metrics.mean_absolute_error method.
Our model achieves an MAE of 0.327.
This means, on average our models predictions are 0.327 units away from the actual value.
Let's make it a little more visual.
You can see the predictions are slightly different to the actual values.
Depending what problem you're working on, having a difference like we do now, might be okay. On the flip side, it may also not be okay, meaning the predictions would have to be closer.
Mean Squared Error (MSE)
How about MSE?
We can calculate it with Scikit-Learn's sklearn.metrics.mean_squared_error.
MSE will often be higher than MAE because is squares the errors rather than only taking the absolute difference into account.
Now you might be thinking, which regression evaluation metric should you use?
R^2 is similar to accuracy. It gives you a quick indication of how well your model might be doing. Generally, the closer your R^2 value is to 1.0, the better the model. But it doesn't really tell exactly how wrong your model is in terms of how far off each prediction is.
MAE gives a better indication of how far off each of your model's predictions are on average.
As for MAE or MSE, because of the way MSE is calculated, squaring the differences between predicted values and actual values, it amplifies larger differences. Let's say we're predicting the value of houses (which we are).
Pay more attention to MAE: When being $10,000 off is twice as bad as being $5,000 off.
Pay more attention to MSE: When being $10,000 off is more than twice as bad as being $5,000 off.
Note: What we've covered here is only a handful of potential metrics you can use to evaluate your models. If you're after a complete list, check out the Scikit-Learn metrics and scoring documentation.
4.2.3 Evaluating a model using the scoring parameter
We've covered a bunch of ways to evaluate a model's predictions but haven't even touched the scoring parameter...
Not to worry, it's very similar to what we've been doing!
As a refresh, the scoring parameter can be used with a function like cross_val_score() to tell Scikit-Learn what evaluation metric to return using cross-validation.
Let's check it out with our classification model and the heart disease dataset.
First, we'll use the default, which is mean accuracy.
We've seen this before, now we got 5 different accuracy scores on different test splits of the data.
Averaging this gives the cross-validated accuracy.
We can find the same using the scoring parameter and passing it "accuracy".
The same goes for the other metrics we've been using for classification.
Let's try "precision".
How about "recall"?
And "f1" (for F1 score)?
We can repeat this process with our regression metrics.
Let's revisit our regression model.
The default is "r2".
Note: We can time how long a single cell of code takes to run using the
%%timemagic command.
But we can use "neg_mean_absolute_error" for MAE (mean absolute error).
Why the "neg_"?
Because Scikit-Learn documentation states:
Which in this case, means a lower negative value (closer to 0) is better.
What about "neg_mean_squared_error" for MSE (mean squared error)?
4.3 Using different evaluation metrics with Scikit-Learn
Remember the third way of evaluating Scikit-Learn functions?
Problem-specific metric functions. Similar to how the
scoringparameter can be passed different scoring functions, Scikit-Learn implements these as stand alone functions.
Well, we've kind of covered this third way of using evaulation metrics with Scikit-Learn.
In essence, all of the metrics we've seen previously have their own function in Scikit-Learn.
They all work by comparing an array of predictions, usually called y_preds to an array of actual labels, usually called y_test or y_true.
Classification functions
For:
Accuracy we can use
sklearn.metrics.accuracy_scorePrecision we can use
sklearn.metrics.precision_scoreRecall we can use
sklearn.metrics.recall_scoreF1 we can use
sklearn.metrics.f1_score
Regression metrics
We can use a similar setup for our regression problem, just with different methods.
For:
R^2 we can use
sklearn.metrics.r2_scoreMAE (mean absolute error) we can use
sklearn.metrics.mean_absolute_errorMSE (mean squared error) we can use
sklearn.metrics.mean_squared_error
Wow!
We've covered a lot!
But it's worth it.
Because evaluating a model's predictions is as important as training a model in any machine learning project.
There's nothing worse than training a machine learning model and optimizing for the wrong evaluation metric.
Keep the metrics and evaluation methods we've gone through when training your future models.
If you're after extra reading, I'd go through the Scikit-Learn guide for model evaluation.
Now we've seen some different metrics we can use to evaluate a model, let's see some ways we can improve those metrics.
5. Improving model predictions through experimentation (hyperparameter tuning)
The first predictions you make with a model are generally referred to as baseline predictions.
It's similar for the first evaluation metrics you get. These are generally referred to as baseline metrics.
Your next goal is to improve upon these baseline metrics.
How?
Experiment, experiment, experiment!
Two of the main methods to improve baseline metrics are:
From a data perspective.
From a model perspective.
From a data perspective asks:
Could we collect more data? In machine learning, more data is generally better, as it gives a model more opportunities to learn patterns.
Could we improve our data? This could mean filling in misisng values or finding a better encoding (turning data into numbers) strategy.
From a model perspective asks:
Is there a better model we could use? If you've started out with a simple model, could you use a more complex one? (we saw an example of this when looking at the Scikit-Learn machine learning map, ensemble methods are generally considered more complex models)
Could we improve the current model? If the model you're using performs well straight out of the box, can the hyperparameters be tuned to make it even better?
Note: Patterns in data are also often referred to as data parameters. The difference between parameters and hyperparameters is a machine learning model seeks to find parameters in data on its own, where as, hyperparameters are settings on a model which a person (you) can adjust.
Since we have two existing datasets, we'll look at improving our results from a model perspective.
More specifically, we'll look at how we could improve our RandomForestClassifier and RandomForestRegressor models through hyperparameter tuning.
What even are hyperparameters?
Good question, let's check them out.
First, we'll instantiate a RandomForestClassifier.
When we instantiate a model like above, we're using the default hyperparameters.
These get printed out when you call the model instance and get_params().
You'll see things like max_depth, min_samples_split, n_estimators.
Each of these is a hyperparameter of the RandomForestClassifier you can adjust.
You can think of hyperparameters as being similar to dials on an oven.
On the default setting your oven might do an okay job cooking your favourite meal. But with a little experimentation, you find it does better when you adjust the settings.
The same goes for imporving a machine learning model by hyperparameter tuning.
The default hyperparameters on a machine learning model may find patterns in data well. But there's a chance a adjusting the hyperparameters may improve a models performance.
Every machine learning model will have different hyperparameters you can tune.
You might be thinking, "how the hell do I remember all of these?"
Another good question.
It's why we're focused on the Random Forest.
Instead of memorizing all of the hyperparameters for every model, we'll see how it's done with one.
And then knowing these principles, you can apply them to a different model if needed.
Reading the Scikit-Learn documentation for the Random Forest, you'll find they suggest trying to change n_estimators (the number of trees in the forest) and min_samples_split (the minimum number of samples required to split an internal node).
We'll try tuning these as well as:
max_features(the number of features to consider when looking for the best split)max_depth(the maximum depth of the tree)min_samples_leaf(the minimum number of samples required to be at a leaf node)
If this still sounds like a lot, the good news is, the process we're taking with the Random Forest and tuning its hyperparameters, can be used for other machine learning models in Scikit-Learn. The only difference is, with a different model, the hyperparameters you tune will be different.
Adjusting hyperparameters is usually an experimental process to figure out which are best. As there's no real way of knowing which hyperparameters will be best when starting out.
To get familar with hyparameter tuning, we'll take our RandomForestClassifier and adjust its hyperparameters in 3 ways.
By hand
Randomly with
sklearn.model_selection.RandomizedSearchCVExhaustively with
sklearn.model_selection.GridSearchCV
5.1 Tuning hyperparameters by hand
So far we've worked with training and test datasets.
You train a model on a training set and evaluate it on a test dataset.
But hyperparameter tuning introduces a thrid set, a validation set.
Now the process becomes:
Train a model on the training data.
(Try to) improve the model's hyperparameters on the validation set.
Evaluate the model on the test set.
If our starting dataset contained 100 different patient records labels indicating who had heart disease and who didn't and we wanted to build a machine learning model to predict who had heart disease and who didn't, it might look like this:
Since we know we're using a RandomForestClassifier and we know the hyperparameters we want to adjust, let's see what it looks like.
First, let's remind ourselves of the base parameters.
And we're going to adjust:
max_depthmax_featuresmin_samples_leafmin_samples_splitn_estimators
We'll use the same code as before, except this time we'll create a training, validation and test split.
With the training set containing 70% of the data and the validation and test sets each containing 15%.
Let's get some baseline results, then we'll tune the model.
And since we're going to be evaluating a few models, let's make an evaluation function.
Wonderful!
Now let's recreate a previous workflow, except we'll add in the creation of a validation set.
Beautiful, now let's try and improve the results.
We'll change 1 of the hyperparameters, n_estimators=100 (default) to n_estimators=200 and see if it improves on the validation set.
Hmm, it looks like doubling the n_estimators value performs worse than the default, perhaps there's a better value for n_estimators?
And what other hyperparameters could we change?
Wait...
This could take a while if all we're doing is building new models with new hyperparameters each time.
Surely there's a better way?
There is.
5.2 Hyperparameter tuning with RandomizedSearchCV
Scikit-Learn's sklearn.model_selection.RandomizedSearchCV allows us to randomly search across different hyperparameters to see which work best.
It also stores details about the ones which work best!
Let's see it in action.
First, we create a dictionary of parameter distributions (collections of different values for specific hyperparamters) we'd like to search over.
This dictionary comes in the form:
Where "hyperparameter_name" is the value of a specific hyperparameter for a model and [values_to_randomly_try] is a list of values for that specific hyperparamter to randomly try.
Where did these values come from?
They're made up.
Made up?
Yes.
Not completely pulled out of the air but after reading the Scikit-Learn documentation on Random Forest's you'll see some of these values have certain values which usually perform well and certain hyperparameters take strings rather than integers.
Now we've got the parameter distribution dictionary setup, Scikit-Learn's RandomizedSearchCV will look at it, pick a random value from each, instantiate a model with those values and test each model.
How many models will it test?
As many as there are for each combination of hyperparameters to be tested. Let's add them up.
Woah!
That's a lot of combinations!
Or...
We can set the n_iter parameter to limit the number of models RandomizedSearchCV tests (e.g. n_iter=20 means to try 20 different random combintations of hyperparameters and will cross-validate each set, so if cv=5, 5x20 = 100 total fits).
The best thing?
The results we get will be cross-validated (hence the CV in RandomizedSearchCV) so we can use train_test_split().
And since we're going over so many different models, we'll set n_jobs=-1 in our RandomForestClassifier so Scikit-Learn takes advantage of all the cores (processors) on our computers.
Let's see it in action.
Note: Depending on
n_iter(how many models you test), the different values in the hyperparameter grid, and the power of your computer, running the cell below may take a while (for reference, it took about ~1-minute on my M1 Pro MacBook Pro).
When RandomizedSearchCV goes through n_iter combinations of of hyperparameter search space, it stores the best ones in the attribute best_params_.
Now when we call predict() on rs_clf (our RandomizedSearchCV version of our classifier), it'll use the best hyperparameters it found.
Excellent!
Thanks to RandomizedSearchCV testing out a bunch of different hyperparameters, we get a nice boost to all of the evaluation metrics for our classification model.
5.3 Hyperparameter tuning with GridSearchCV
There's one more way we could try to improve our model's hyperparamters.
And it's with sklearn.model_selection.GridSearchCV.
The main difference between GridSearchCV and RandomizedSearchCV is GridSearchCV searches across a grid of hyperparamters exhaustively (it will try every combination possible), where as, RandomizedSearchCV searches across a grid of hyperparameters randomly (stopping after n_iter combinations).
GridSearchCV also refers to a dictionary of parameter distributions as a parameter grid (via the parameter param_grid).
For example, let's see our dictionary of hyperparameters.
RandomizedSearchCV tries n_iter combinations of different values.
Where as, GridSearchCV will try every single possible combination.
And if you remember from before when we did the calculation: max_depth has 4 values, max_features has 2, min_samples_leaf has 3, min_samples_split has 3, n_estimators has 5.
That's 4x2x3x3x5 = 360 models!
This could take a long time depending on the power of the computer you're using, the amount of data you have and the complexity of the hyperparamters (usually higher values means a more complex model).
In our case, the data we're using is relatively small (only ~300 samples).
Since we've already tried to find some ideal hyperparameters using RandomizedSearchCV, we'll create another hyperparameter grid based on the best_params_ of rs_clf with less options and then try to use GridSearchCV to find a more ideal set.
In essence, the workflow could be:
Tune hyperparameters by hand to get a feel of the data/model.
Create a large set of hyperparameter distributions and search across them randomly with
RandomizedSearchCV.Find the best hyperparameters from 2 and reduce the search space before searching across a smaller subset exhaustively with
GridSearchCV.
Note: Based on the
best_params_ofrs_clfimplies the next set of hyperparameters we'll try are roughly in the same range of the best set found byRandomizedSearchCV.
We've created another grid of hyperparameters to search over, this time with less total.
Now when we run GridSearchCV, passing it our classifier (clf), parameter grid (param_grid) and the number of cross-validation folds we'd like to use (cv=5), it'll create a model with every single combination of hyperparameters, and then cross-validate each 5 times (for example, 36 hyperparameter combinations * 5 = 135 fits in total) and check the results.
Note: Depending on the compute power of the machine you're using, the following cell may take a few minutes to run (for reference, it took ~60 seconds on my M1 Pro MacBook Pro).
Once it completes, we can check the best hyperparameter combinations it found using the best_params_ attribute.
And by default when we call the predict() function on gs_clf, it'll use the best hyperparameters.
Let's create a DataFrame to compare the different metrics.
Nice!
After trying many different combinations of hyperparamters, we get a slight improvement in results.
However, sometimes you'll notice that your results don't change much.
These things might happen.
But it's important to remember, it's not over. There more things you can try.
In a hyperparameter tuning sense, there may be a better set we could find through more extensive searching with RandomizedSearchCV and GridSearchCV, this would require more experimentation.
Other techniques you could:
Collecting more data - Based on the results our models are getting now, it seems like they're very capable of finding patterns. Collecting more data may improve a models ability to find patterns. However, your ability to do this will largely depend on the project you're working on.
Try a more advanced model - Although our tuned Random Forest model is doing pretty well, a more advanced ensemble method such as XGBoost or CatBoost might perform better. I'll leave these for extra-curriculum.
Since machine learning is part engineering, part science, these kind of experiments are common place in any machine learning project.
Now we've got a tuned Random Forest model, let's find out how we might save it and export it so we can share it with others or potentially use it in an external application.
6. Saving and loading trained machine learning models
Our GridSearchCV model (gs_clf) has the best results so far, we'll export it and save it to file.
6.1 Saving and loading a model with pickle
We saw right at the start, one way to save a model is using Python's pickle module.
We'll use pickle's dump() method and pass it our model, gs_clf, along with the open() function containing a string for the filename we want to save our model as, along with the "wb" string which stands for "write binary", which is the file type open() will write our model as.
Once it's saved, we can import it using pickle's load() function, passing it open() containing the filename as a string and "rb" standing for "read binary".
Once you've reimported your trained model using pickle, you can use it to make predictions as usual.
You'll notice the reimported model evaluation metrics are the same as the model before we exported it.
6.2 Saving and loading a model with joblib
The other way to load and save models is with joblib. Which works relatively the same as pickle.
To save a model, we can use joblib's dump() function, passing it the model (gs_clf) and the desired filename.
Once you've saved a model using dump(), you can import it using load() and passing it the filename of the model.
Again, once imported, we can make predictions with our model.
And once again, you'll notice the evaluation metrics are the same as before.
So which one should you use, pickle or joblib?
According to Scikit-Learn's model persistence documentation, they suggest it may be more efficient to use joblib as it's more efficient with large numpy arrays (which is what may be contained in trained/fitted Scikit-Learn models).
7. Revisiting the entire pipeline
We've covered a lot. And so far, it seems to be all over the place, which it is.
But not to worry, machine learning projects often start out like this.
A whole bunch of experimenting and code all over the place at the start and then once you've found something which works, the refinement process begins.
What would this refinement process look like?
We'll use the car sales regression problem (predicting the sale price of cars) as an example.
To tidy things up, we'll be using Scikit-Learn's sklearn.pipeline.Pipeline class.
You can imagine Pipeline as being a way to string a number of different Scikit-Learn processes together.
7.1 Creating a regression Pipeline
You might recall when, way back in Section 2: Getting Data Ready, we dealt with the car sales data, to build a regression model on it, we had to encode the categorical features into numbers and fill the missing data.
The code we used worked, but it was a bit all over the place.
Good news is, Pipeline can help us clean it up.
Let's remind ourselves what our car-sales-extended-missing-data.csv looks like in DataFrame form.
There's 1000 rows, three features are categorical (Make, Colour, Doors), the other two are numerical (Odometer (KM), Price) and there's 249 missing values.
We're going to have to turn the categorical features into numbers and fill the missing values before we can fit a model.
We'll build a Pipeline to do so.
Pipeline's main input parameter is steps which is a list of tuples ([(step_name, action_to_take)]) of the step name, plus the action you'd like it to perform.
In our case, you could think of the steps as:
Fill missing data
Convert data to numbers
Build a model on the data
Let's do it!
What we've done is combine a series of data preprocessing steps (filling missing values, encoding numerical values) as well as a model into a Pipeline.
Doing so not only cleans up the code, it ensures the same steps are taken every time the code is run rather than having multiple different processing steps happening in different stages.
It's also possible to GridSearchCV or RandomizedSearchCV with a Pipeline.
The main difference is when creating a hyperparameter grid, you have to add a prefix to each hyperparameter (see the documentation for RandomForestRegressor for a full list of possible hyperparameters to tune).
The prefix is the name of the Pipeline step you'd like to alter, followed by two underscores.
For example, to adjust n_estimators of "model" in the Pipeline, you'd use: "model__n_estimators" (note the double underscore after model__ at the start).
Let's see it!
Note: Depending on your computer's processing power, the cell below may take a few minutes to run. For reference, it took about ~60 seconds on my M1 Pro MacBook Pro.
Now let's find the score of our model (by default GridSearchCV saves the best model to the gs_model object).
Beautiful!
Using GridSearchCV we see a nice boost in our models score.
And the best thing is, because it's all in a Pipeline, we could easily replicate these results.
Where to next?
If you've made it this far, congratulations! We've covered a lot of ground in the Scikit-Learn library.
As you might've guessed, there's a lot more to be discovered.
But for the time being, you should be equipped with some of the most useful features of the library to start trying to apply them to your own problems.
Somewhere you might like to look next is to apply what you've learned above to a Kaggle competition.
Kaggle competitions are great places to practice your data science and machine learning skills and compare your results with others.
A great idea would be to try to combine the heart disease classification code, as well as the Pipeline code, to build a model for the Titanic dataset.
Otherwise, if you'd like to figure out what else the Scikit-Learn library is capable of I'd highly recommend browsing through the Scikit-Learn User Guide and seeing what sparks your interest.
Finally, as an extra-curriculum extension, you might want to look into trying out the CatBoost library for dealing with non-numerical data automatically.
The CatBoost algorithm is advanced version of a decision tree (like a Random Forest with superpowers) and is used in production at several large technology companies, including Cloudflare.