ENH Rework narrative of GBDT notebook

Author: ArturoAmorQ

python_scripts/ensemble_gradient_boosting.py

@@ -8,26 +8,28 @@
 # %% [markdown]
 # # Gradient-boosting decision tree (GBDT)
 #
-# In this notebook, we will present the gradient boosting decision tree
-# algorithm and contrast it with AdaBoost.
+# In this notebook, we present the gradient boosting decision tree algorithm.
 #
-# Gradient-boosting differs from AdaBoost due to the following reason: instead
-# of assigning weights to specific samples, GBDT will fit a decision tree on the
-# residuals error (hence the name "gradient") of the previous tree. Therefore,
-# each new tree in the ensemble predicts the error made by the previous learner
-# instead of predicting the target directly.
+# Even if AdaBoost and GBDT are both boosting algorithms, they are different in
+# nature: the former assigns weights to specific samples, whereas GBDT fits
+# succesive decision trees on the residual errors (hence the name "gradient") of
+# their preceding tree. Therefore, each new tree in the ensemble tries to refine
+# its predictions by specifically addressing the errors made by the previous
+# learner, instead of predicting the target directly.
 #
-# In this section, we will provide some intuition about the way learners are
-# combined to give the final prediction. In this regard, let's go back to our
-# regression problem which is more intuitive for demonstrating the underlying
+# In this section, we provide some intuitions on the way learners are combined
+# to give the final prediction. For such purpose, we tackle a single-feature
+# regression problem, which is more intuitive for demonstrating the underlying
 # machinery.
+#
+# Later in this notebook we compare the performance of GBDT (boosting) with that
+# of a Random Forest (bagging) for a particular dataset.
 
 # %%
 import pandas as pd
 import numpy as np
 
-# Create a random number generator that will be used to set the randomness
-rng = np.random.RandomState(0)
+rng = np.random.RandomState(0)  # Create a random number generator
 
 
 def generate_data(n_samples=50):
@@ -60,9 +62,9 @@ def generate_data(n_samples=50):
 _ = plt.title("Synthetic regression dataset")
 
 # %% [markdown]
-# As we previously discussed, boosting will be based on assembling a sequence of
-# learners. We will start by creating a decision tree regressor. We will set the
-# depth of the tree so that the resulting learner will underfit the data.
+# As we previously discussed, boosting is based on assembling a sequence of
+# learners. We start by creating a decision tree regressor. We set the depth of
+# the tree to underfit the data on purpose.
 
 # %%
 from sklearn.tree import DecisionTreeRegressor
@@ -74,45 +76,61 @@ def generate_data(n_samples=50):
 target_test_predicted = tree.predict(data_test)
 
 # %% [markdown]
-# Using the term "test" here refers to data that was not used for training. It
-# should not be confused with data coming from a train-test split, as it was
-# generated in equally-spaced intervals for the visual evaluation of the
-# predictions.
+# Using the term "test" here refers to data not used for training. It should not
+# be confused with data coming from a train-test split, as it was generated in
+# equally-spaced intervals for the visual evaluation of the predictions.
+#
+# To avoid writing the same code in multiple places we define a helper function
+# to plot the data samples as well as the decision tree predictions and
+# residuals.
+
 
 # %%
-# plot the data
-sns.scatterplot(
-    x=data_train["Feature"], y=target_train, color="black", alpha=0.5
-)
-# plot the predictions
-line_predictions = plt.plot(data_test["Feature"], target_test_predicted, "--")
+def plot_decision_tree_with_residuals(y_train, y_train_pred, y_test_pred):
+    # Create a plot and get the Axes object
+    fig, ax = plt.subplots()
+    # plot the data
+    sns.scatterplot(
+        x=data_train["Feature"], y=y_train, color="black", alpha=0.5, ax=ax
+    )
+    # plot the predictions
+    line_predictions = ax.plot(data_test["Feature"], y_test_pred, "--")
+
+    # plot the residuals
+    for value, true, predicted in zip(
+        data_train["Feature"], y_train, y_train_pred
+    ):
+        lines_residuals = ax.plot(
+            [value, value], [true, predicted], color="red"
+        )
+
+    handles = [line_predictions[0], lines_residuals[0]]
+
+    return handles, ax
 
-# plot the residuals
-for value, true, predicted in zip(
-    data_train["Feature"], target_train, target_train_predicted
-):
-    lines_residuals = plt.plot([value, value], [true, predicted], color="red")
 
-plt.legend(
-    [line_predictions[0], lines_residuals[0]], ["Fitted tree", "Residuals"]
+handles, ax = plot_decision_tree_with_residuals(
+    target_train, target_train_predicted, target_test_predicted
 )
-_ = plt.title("Prediction function together \nwith errors on the training set")
+legend_labels = ["Initial decision tree", "Initial residuals"]
+ax.legend(handles, legend_labels, bbox_to_anchor=(1.05, 0.8), loc="upper left")
+_ = ax.set_title("Decision Tree together \nwith errors on the training set")
 
 # %% [markdown]
 # ```{tip}
 # In the cell above, we manually edited the legend to get only a single label
 # for all the residual lines.
 # ```
 # Since the tree underfits the data, its accuracy is far from perfect on the
-# training data. We can observe this in the figure by looking at the difference
-# between the predictions and the ground-truth data. We represent these errors,
-# called "Residuals", by unbroken red lines.
+# training data. We can observe this in the figure above by looking at the
+# difference between the predictions and the ground-truth data. We represent
+# these errors, called "residuals", using solid red lines.
 #
-# Indeed, our initial tree was not expressive enough to handle the complexity of
+# Indeed, our initial tree is not expressive enough to handle the complexity of
 # the data, as shown by the residuals. In a gradient-boosting algorithm, the
-# idea is to create a second tree which, given the same data `data`, will try to
-# predict the residuals instead of the vector `target`. We would therefore have
-# a tree that is able to predict the errors made by the initial tree.
+# idea is to create a second tree which, given the same `data`, tries to predict
+# the residuals instead of the vector `target`, i.e. we have a second tree that
+# is able to predict the errors made by the initial tree.
 #
 # Let's train such a tree.
 
@@ -126,29 +144,22 @@ def generate_data(n_samples=50):
 target_test_predicted_residuals = tree_residuals.predict(data_test)
 
 # %%
-sns.scatterplot(x=data_train["Feature"], y=residuals, color="black", alpha=0.5)
-line_predictions = plt.plot(
-    data_test["Feature"], target_test_predicted_residuals, "--"
+handles, ax = plot_decision_tree_with_residuals(
+    residuals,
+    target_train_predicted_residuals,
+    target_test_predicted_residuals,
 )
-
-# plot the residuals of the predicted residuals
-for value, true, predicted in zip(
-    data_train["Feature"], residuals, target_train_predicted_residuals
-):
-    lines_residuals = plt.plot([value, value], [true, predicted], color="red")
-
-plt.legend(
-    [line_predictions[0], lines_residuals[0]],
-    ["Fitted tree", "Residuals"],
-    bbox_to_anchor=(1.05, 0.8),
-    loc="upper left",
-)
-_ = plt.title("Prediction of the previous residuals")
+legend_labels = [
+    "Predicted residuals",
+    "Residuals of the\npredicted residuals",
+]
+ax.legend(handles, legend_labels, bbox_to_anchor=(1.05, 0.8), loc="upper left")
+_ = ax.set_title("Prediction of the initial residuals")
 
 # %% [markdown]
-# We see that this new tree only manages to fit some of the residuals. We will
-# focus on a specific sample from the training set (i.e. we know that the sample
-# will be well predicted using two successive trees). We will use this sample to
+# We see that this new tree only manages to fit some of the residuals. We now
+# focus on a specific sample from the training set (as we know that the sample
+# can be well predicted using two successive trees). We will use this sample to
 # explain how the predictions of both trees are combined. Let's first select
 # this sample in `data_train`.
 
@@ -159,66 +170,49 @@ def generate_data(n_samples=50):
 target_true_residual = residuals.iloc[-2]
 
 # %% [markdown]
-# Let's plot the previous information and highlight our sample of interest.
-# Let's start by plotting the original data and the prediction of the first
-# decision tree.
+# Let's plot the original data, the predictions of the initial decision tree and
+# highlight our sample of interest, i.e. this is just a zoom of the plot
+# displaying the initial shallow tree.
 
 # %%
-# Plot the previous information:
-#   * the dataset
-#   * the predictions
-#   * the residuals
-
-sns.scatterplot(
-    x=data_train["Feature"], y=target_train, color="black", alpha=0.5
+handles, ax = plot_decision_tree_with_residuals(
+    target_train, target_train_predicted, target_test_predicted
 )
-plt.plot(data_test["Feature"], target_test_predicted, "--")
-for value, true, predicted in zip(
-    data_train["Feature"], target_train, target_train_predicted
-):
-    lines_residuals = plt.plot([value, value], [true, predicted], color="red")
-
-# Highlight the sample of interest
-plt.scatter(
+ax.scatter(
     sample, target_true, label="Sample of interest", color="tab:orange", s=200
 )
-plt.xlim([-1, 0])
-plt.legend(bbox_to_anchor=(1.05, 0.8), loc="upper left")
-_ = plt.title("Tree predictions")
+ax.set_xlim([-1, 0])
+ax.legend(bbox_to_anchor=(1.05, 0.8), loc="upper left")
+_ = ax.set_title("Zoom of sample of interest\nin the initial decision tree")
 
 # %% [markdown]
-# Now, let's plot the residuals information. We will plot the residuals computed
-# from the first decision tree and show the residual predictions.
+# Similarly we plot a zoom of the plot with the prediction of the initial residuals
 
 # %%
-# Plot the previous information:
-#   * the residuals committed by the first tree
-#   * the residual predictions
-#   * the residuals of the residual predictions
-
-sns.scatterplot(x=data_train["Feature"], y=residuals, color="black", alpha=0.5)
-plt.plot(data_test["Feature"], target_test_predicted_residuals, "--")
-for value, true, predicted in zip(
-    data_train["Feature"], residuals, target_train_predicted_residuals
-):
-    lines_residuals = plt.plot([value, value], [true, predicted], color="red")
-
-# Highlight the sample of interest
+handles, ax = plot_decision_tree_with_residuals(
+    residuals,
+    target_train_predicted_residuals,
+    target_test_predicted_residuals,
+)
 plt.scatter(
     sample,
     target_true_residual,
     label="Sample of interest",
     color="tab:orange",
     s=200,
 )
-plt.xlim([-1, 0])
-plt.legend()
-_ = plt.title("Prediction of the residuals")
+legend_labels = [
+    "Predicted residuals",
+    "Residuals of the\npredicted residuals",
+]
+ax.set_xlim([-1, 0])
+ax.legend(bbox_to_anchor=(1.05, 0.8), loc="upper left")
+_ = ax.set_title("Zoom of sample of interest\nin the initial residuals")
 
 # %% [markdown]
 # For our sample of interest, our initial tree is making an error (small
 # residual). When fitting the second tree, the residual in this case is
-# perfectly fitted and predicted. We will quantitatively check this prediction
+# perfectly fitted and predicted. We can quantitatively check this prediction
 # using the fitted tree. First, let's check the prediction of the initial tree
 # and compare it with the true value.
 
@@ -265,7 +259,9 @@ def generate_data(n_samples=50):
 # second tree corrects the first tree's error, while the third tree corrects the
 # second tree's error and so on).
 #
-# We will compare the generalization performance of random-forest and gradient
+# ## First comparison of GBDT vs random forests
+#
+# We now compare the generalization performance of random-forest and gradient
 # boosting on the California housing dataset.
 
 # %%
@@ -322,11 +318,12 @@ def generate_data(n_samples=50):
 print(f"Average score time: {cv_results_rf['score_time'].mean():.3f} seconds")
 
 # %% [markdown]
-# In term of computation performance, the forest can be parallelized and will
+# In terms of computing performance, the forest can be parallelized and then
 # benefit from using multiple cores of the CPU. In terms of scoring performance,
 # both algorithms lead to very close results.
 #
-# However, we see that the gradient boosting is a very fast algorithm to predict
-# compared to random forest. This is due to the fact that gradient boosting uses
-# shallow trees. We will go into details in the next notebook about the
-# hyperparameters to consider when optimizing ensemble methods.
+# However, we see that gradient boosting is overall faster than random forest.
+# One of the reasons is that random forests typically rely on deep trees (that
+# overfit individually) whereas boosting models build shallow trees (that
+# underfit individually) which are faster to fit and predict. In the following
+# exercise we will explore more in depth how these two models compare.