Chapter 6 - Decision Trees¶
"Decision Trees are versatile Machine Learning algorithms that can perform both classification and regression tasks, and even multioutput tasks." [Geron2017]
%matplotlib inline
import numpy as np
np.random.seed(42) # to ensure our results exactly like the book
6.1 Training and Visualizing a Decision Tree¶
First, let's load the iris dataset from sci-kit learn library.
from sklearn.datasets import load_iris
from sklearn.tree import DecisionTreeClassifier
iris = load_iris()
6.1.1 Determine Targets¶
Let's determine which columns will be in our interest and print them.
X = iris.data[:, 2:] # only focus on petal length and width
Y = iris.target
feature_names = iris.feature_names[2:]
print("given:",feature_names,
"\npredict whether:", iris.target_names)
Exercise 6.1.1: Plot the data set¶
Plot the data set and have a look at the two features that are selected.
# use matplotlib as you did on previous labs
6.1.2 Train the dataset¶
Without separating the dataset as we did in previous labs, let's use all the data set and train the decision tree.
tree_clf = DecisionTreeClassifier(max_depth=2)
tree_clf.fit(X,Y)
There are many hyperparameters that a decision tree classifier has. You can see from the ouput above the parameters that will be used for predictions. Two criterions that you can use with decision trees in sci-kit learn. These metrics are calculated in each node of decision tree.
- Gini impurity
criterion='gini'is a measure of how often a randomly chosen element from a set would be incorrectly labeled. Formally it is computed by: $$ I_G(p) = \sum_{i=1}^{J} p_i \sum_{k \neq i} p_k $$ where $J$ denotes classes and $p_i$ is the fraction of items which are labeled with class $i$. For a concrete example have a look: https://stats.stackexchange.com/a/339514 - Information Gain
criterion='entropy'is a mesaure of entropy, which is used in thermodynamics as a measure of molecular disorder. Entropy=0 means the molecules are well ordered.
$$ H(T) = I_E(p)=-\sum_{i=1}^{J}p_i\log_2 p_i $$ $p_1, p_2, ...$ as in gini impurity are the fractions that add up to 1.
These two metrics are used for deciding the splits while training a decision tree.
Exercise 6.1.2: Gini or Entropy¶
- Should you use Gini impurity or entropy?
- Which one is faster to compute and why?
- Visualize the tree that you trained above in 6.1.2.
- Why does gini impurity metric decrease in deeper nodes?
- Which cases do you see that the metric is zero?
Your answer here
6.1.2 Visualization¶
You can export the decision tree as a dot file from sci-kit learn. You can convert dot to png image by installing graphviz.
from sklearn.tree import export_graphviz
export_graphviz(tree_clf,
out_file="iris_tree.dot",
feature_names=feature_names,
class_names=iris.target_names,
rounded=True,
filled=True
)
# Make sure you installed graphviz (exclamation mark is for shell commands)
!apt install graphviz
# Convert dot file to png file.
!dot -Tpng iris_tree.dot -o iris_tree.png
from IPython.display import Image
Image(filename='iris_tree.png')
To see a better visualization example of decision trees, have a look at this page.
Extra: Another way to visualize decision trees¶
There is a brand new visualization library from creators of ANTLR (parser generator) for decision trees called dtreeviz. You can find some other examples from their repository for better visualization. Follow the steps below:
# install the package
!pip install dtreeviz
# (optional)
!apt-get install msttcorefonts -qq
from dtreeviz.trees import dtreeviz
import matplotlib as mpl
mpl.rcParams['axes.facecolor'] = 'white'
viz = dtreeviz(tree_clf,
X,
Y,
target_name='flower type',
feature_names=feature_names,
class_names=list(iris.target_names),
fancy=True,
orientation ='TD')
# uncomment this
# viz
6.1.3 Plot training data set¶
import matplotlib.pyplot as plt
from matplotlib.colors import ListedColormap
CUSTOM_CMAP = ListedColormap(['#fafab0','#9898ff','#a0faa0'])
# helper function to plot the boundaries
def plot_decision_boundary(clf, x, y):
color_map = ["yo", "bs", "g^"]
for target_index, target_name in enumerate(iris.target_names):
plt.plot(x[:, 0][y==target_index], # petal length on X axis (the ones that equal to target)
x[:, 1][y==target_index], # petal width on Y axis (the ones that equal to target)
color_map[target_index],
label=target_name)
x1s = np.linspace(np.min(x[:, 0]), np.max(x[:, 0]), 100)
x2s = np.linspace(np.min(x[:, 1]), np.max(x[:, 1]), 100)
x1, x2 = np.meshgrid(x1s, x2s)
x_test = np.c_[x1.ravel(), x2.ravel()]
y_pred = clf.predict(x_test).reshape(x1.shape)
plt.contourf(x1, x2, y_pred, alpha=0.3, cmap=CUSTOM_CMAP)
plot_decision_boundary(tree_clf, X, Y)
plt.xlabel(feature_names[0]) # petal length (cm)
plt.ylabel(feature_names[1]) # petal width (cm)
plt.show()
Exercise 6.1: Plot decision boundaries¶
- Write down which colour correspond to which node in the decision tree.
- Try to increase/decrease
max_depthof decision tree classifier and observe the changes in the decision boundaries. What would you setmax_depth? - If the helper function (
plot_decision_boundary) was not available to you, how would you visualize the decision boundaries? Tip: try to create X's that ranges from $[1, 0]$ to $[7, 2.5]$ where. You can use:
# check np.mgrid[minX1:maxX1:increment, minX2:maxX2:increment]
X = np.mgrid[0:10:1, -5:0:1].reshape(2,-1).T
X
Exercise 6.1a: Solution¶
# Tip: try each point in the space
6.1.4 Estimating Class Probabilities¶
To estimate the probability of an instance belongs to a class, you can use predict_proba, to determine the class that an instance will be assigned to use predict.
tree_clf.predict_proba([[5, 1.5]])
tree_clf.predict([[5, 1.5]])
6.1.5 Regularization Hyperparameters¶
"Constraining a model to make it simpler and reduce the risk of overfitting is called _regularization" [Geron2017, page 27] To avoid overfitting, you can limit the generation of a node by min_samples_leaf (the minimum samples that a node must have to able to be splitted.).
from sklearn.datasets import make_moons
Xm, ym = make_moons(n_samples=100, noise=0.25, random_state=53)
deep_tree_clf1 = DecisionTreeClassifier(random_state=42)
deep_tree_clf2 = DecisionTreeClassifier(min_samples_leaf=4, random_state=42)
deep_tree_clf1.fit(Xm, ym)
deep_tree_clf2.fit(Xm, ym)
plt.figure(figsize=(11, 4))
plt.subplot(121)
plt.xlabel(r"$x_1$", fontsize=18)
plt.ylabel(r"$x_2$", fontsize=18, rotation=0)
plot_decision_boundary(deep_tree_clf1, Xm, ym)
plt.title("No restrictions", fontsize=16)
plt.subplot(122)
plt.xlabel(r"$x_1$", fontsize=18)
plt.ylabel(r"$x_2$", fontsize=18, rotation=0)
plot_decision_boundary(deep_tree_clf2, Xm, ym)
plt.title("min_samples_leaf = {}".format(deep_tree_clf2.min_samples_leaf), fontsize=14)
plt.show()
6.2 Regression¶
Decision trees can be used for regression tasks too. Instead of predicting a class, in regression tasks, the aim is to predict a numberic value (such as the price of a car). Assume that we have this quadratic data set with some noise:
# Quadratic training set + noise
np.random.seed(42)
m = 200
X = np.random.rand(m, 1)
y = 4 * (X - 0.5) ** 2
Y = y + np.random.randn(m, 1) / 10
plt.plot(X, Y, "bo")
plt.xlabel("$x_{1}$", fontsize=18)
plt.ylabel("$y$", fontsize=18, rotation=0)
plt.show()
from sklearn.tree import DecisionTreeRegressor
tree_reg = DecisionTreeRegressor(max_depth=2)
tree_reg.fit(X,Y)
Exercise 6.2: Visualize the regression tree¶
- Visualize the regression tree same as before with graphviz.
- Plot this regression tree (Tip: try many values for x (e.g.
np.linspace(min, max, noOfPoints)) - Plot the decision boundaries of
max_depth=2andmax_depth=3regression trees (also trymin_samples_leaf=10) - Compare the differences on the difference plots. Notice that the average is taken in the regions which are separated by the decision tree regressor.
# Tips for 6.2.2:
Xs = np.linspace(0, 1, 100)
Xs
# predict Y values for Xs and plot
Chapter 7: Ensemble Learning and Random Forests¶
Instead of using a single predictor, to improve our predictions we use now use an ensemble: a group of predictors. It is as if you are asking a number of experts opinion about a problem and you aggregate their answers.
7.1 Voting Classifiers¶
A brief comparision between soft voting and hard voting with using three predictors.
from sklearn.model_selection import train_test_split
from sklearn.datasets import make_moons
X, Y = make_moons(n_samples=500, noise=0.30, random_state=42)
X_train, X_test, Y_train, Y_test = train_test_split(X, Y, random_state=42)
Helper function for printing accuracies on test set
from sklearn.metrics import accuracy_score
def test_clfs(*clfs): # clf -> classifier
for clf in clfs:
clf.fit(X_train, Y_train) # train the classifier
Y_pred = clf.predict(X_test)
print(clf.__class__.__name__ + ":", accuracy_score(Y_test, Y_pred))
Let's test hard voting first.
from sklearn.ensemble import RandomForestClassifier
from sklearn.ensemble import VotingClassifier
from sklearn.linear_model import LogisticRegression
from sklearn.svm import SVC
# Don't worry about the warnings,
# sci-kit community will be fixing it in the next major version 0.20.0
log_clf = LogisticRegression(random_state=42)
rnd_clf = RandomForestClassifier(random_state=42)
svm_clf = SVC(random_state=42, probability=True)
voting_clf = VotingClassifier(estimators=[('lr', log_clf),
('rf', rnd_clf),
('svc', svm_clf)],
voting='hard')
test_clfs(log_clf, rnd_clf, svm_clf, voting_clf)
Exercise 7.1: Which voting to pick?¶
- Check the
softvoting and compare the results. Why do you think it is different?
If you don't know the difference have a look at your book, page 186.
Your answer here:
7.2 Bagging & Pasting¶
Instead of giving training set to each predictor in our ensemble, another approach to gain more accuracy is to separate the training set and give a different training subset to each predictor. There are two ways:
- Bagging: (bootstrapping in statistics) picking a random subset from training set and not removing this selected set from training set for each predictor, i.e. sampling with replacement
- Pasting: picking a random subset from training set and removing this selected set from the training set for each predictor, i.e. sampling without replacement If you are still confused, have a look at here.
from sklearn.ensemble import BaggingClassifier
from sklearn.tree import DecisionTreeClassifier
# define our decision tree classifier
tree_clf = DecisionTreeClassifier(random_state=42)
# 500 copies of the predictor, which has 100 samples from training set
# n_jobs=-1 for utilizing all cores
bag_clf = BaggingClassifier(tree_clf,
n_estimators=500,
max_samples=100,
bootstrap=True,
n_jobs=-1,
random_state=42)
# fit the bagging classifier
bag_clf.fit(X_train, Y_train)
tree_clf.fit(X_train, Y_train)
Y_pred_bag = bag_clf.predict(X_test)
Y_pred_tree = tree_clf.predict(X_test)
from sklearn.metrics import accuracy_score
print("Bagging Classifier")
print(accuracy_score(Y_test, Y_pred_bag))
print("Decision Tree Classifier")
print(accuracy_score(Y_test, Y_pred_tree))
plt.figure(figsize=(11,4))
plt.subplot(121)
plot_decision_boundary(tree_clf, X, Y)
plt.title("Decision Tree", fontsize=14)
plt.subplot(122)
plot_decision_boundary(bag_clf, X, Y)
plt.title("Decision Trees with Bagging", fontsize=14)
plt.show()
7.2 Exercise: Bagging or Pasting¶
- What is the difference between these two techniques which select a subset of a training set?
- When do you think it is better to use bagging or pasting?
7.3 Gradient Boosting¶
In Gradient Boosting, we start with a single predictor. Sequentially, we add new predictors to an ensemble that corrects its predecessor. In detail:
The algorithm is for 3 predictors in our ensemble:
- Pick a set of weak predictors (e.g. DecisionTreeRegressor(max_depth=2))
- For this example, it is
3.
- For this example, it is
- Train the initial predictor on training set
- Train the second predictor on residual errors (use only errors, i.e.
y2 = y - predictor1.predict(X)) - Train the third predictor on the residual errors of the second predictor
- Use all predictors and sum their predictions (if regression)
Exercise 7.3: Implement, Plot and Compare¶
- Implement this boosting technique that is explained above with
npredictors.- Use the
quadratic training setas a regression task. - Use
DecisionTreeRegressor(max_depth=2)as a weak predictor
- Use the
- For $n=3$, plot step by step the evolution of the ensemble predictions
- First figure with only one predictor
- Second figure with two predictors
- Third figure with three predictors
- Use sci-kit learn's
GradientBoostingRegressorand compare if your results are similiar. - Do you think this boosting technique is scalable in practice?
Tips:
- Plotting: this is similar as before in Regression section 6.2, as explained earlier.
- GradientBoostingRegressor: use this code snippet:
from sklearn.ensemble import GradientBoostingRegressor gbrt = GradientBoostingRegressor(max_depth=2, n_estimators=3, learning_rate=1.0) gbrt.fit(X,Y)
# Your implementation here
Recap¶
At this point, we demonstrated these concepts:
- training and visualizing various models from Decision Trees and Ensemble Learning
- details about some hyperparameters and how to regularize the models
- different training methods for ensemble learning
- how a boosting technique works and its properties.
Of course, there are some material that we have not be able to cover. In your free time, it can be better to have a look at:
- Decision Trees
- How hard is it to find the best split to construct the optimal tree?
- Why is CART algorithm called a greedy algorithm?
- Ensemble Learning and Random Forests
- Difference between Extra-Trees and Random Forests
- How does AdaBoost work?
- Learning Feature Importance with Random Forests
- What is
Stacking?