5.4 Decision Trees

A decision tree is a decision support tool that uses a tree-like model of decisions and their possible consequences, including chance event outcomes, resource costs, and utility. It is one way to display an algorithm that only contains conditional control statements.

Decision trees are commonly used in operations research, specifically in decision analysis, to help identify a strategy most likely to reach a goal, but are also a popular tool in machine learning. — Wikipedia

In Scikit Learn you can also find the Decision Trees.

5.4.1 Scikit Learn Decision Tree

This classifier has the same workflow from the other two (GaussianNB and SVM).

# Import decision tree from sklearn.
from sklearn import tree

# Create the classifier using tree.
clf = tree.DecisionTreeClassifier()

# Fitting/Training the model.
clf.fit(df_train, labels_train)

# Predicting
pred = clf.predict(df_test)

5.4.2 Parameters

The .DecisionTreeClassifier() has a lot of parameters to set up.

5.4.2.1 min_samples_split

This parameter set a limit of how deep the decision tree will go. Figure 1 shows an example of the extension of the Tree.

c5_l4_01.png

Figure 1 - Example of leafs and nodes.


As you can see the min_samples_split will go further for each node until reach the node with two leafs.

The effect of going further and further, it means, a very low number of min_samples_split is more likely prone to overfitting. You can see an example of overfitting in Figure 2.

c5_l4_02.png

Figure 2 - Overfitting.


5.4.2.2 Entropy

Bear in mind, the default function to measure the quality of a split in DecisionTreeClassifier() is the giniindex instead of the entropy.

The entropy controls how the decision tree decides where to split the data.

\[entropy = - \sum_i^n p_i \cdot (\ln(p_i))\]

When entropy is 1 there is no pattern, the data was spread equally on the “plot” and you can not realize where to split. On the other hand, when entropy is 0 there are a clearly areas to split.

Though, the entropy is applied to parents and child of the tree in a way that you will “compare” entropies from parents and child.

5.4.2.3 Gain

Founded on the entropies calculated from the parent and child, it is possible to calculate the gain.

\[\text{Information Gain} = entropy(parent) - [\text{weighted average]} \cdot entropy(children)\]

c5_l4_03.png

Figure 3 - Information Gain.


The object function of a decision tree is to maximize the information gain.

5.4.2.4 Example

Figure 4 shows a dataframe to be splited.

c5_l4_04.png

Figure 4 - How to split using Decision Tree?.


Bear in mind, the parent entropy of the all dataset is 1.

Now, let’s use the gain information to decide which variable to use.

Grade

Ilustrating the decision tree using a tree, Figure 5 shows a summary.

c5_l4_05.png

Figure 5 - The grade splitting.


Splitting by grade flat has only fast, whereas steep has two slow and 1 gast. Let’s calculate the probabilities/proportion in Table 1:

Table 1 - Table of quantity and proportion for Grade
Steep Flat psteep pflat
Slow 2 1 2/3 1/3
Fast 0 1 0/1 1/1

Based on these probabilities, I can calculate the entropy of each category (steep and flat).

  • Entropy calculation:

\[entropy(flat) = -(p_{flat,slow} \cdot \ln(p_{flat,slow}) + p_{flat,fast} \cdot \ln(p_{flat,fast})) = \\ -(1 \cdot \ln(1) + 0 \cdot \ln(0)) = 0 \]

\[entropy(steep) = -(p_{steep,slow} \cdot \ln(p_{steep,slow}) + p_{steep,fast} \cdot \ln(p_{steep,fast})) = -(\frac{2}{3} \cdot \ln(\frac{2}{3}) + \frac{1}{3} \cdot \ln(\frac{1}{3})) = 0.9183\]

  • Information Gain:

\[\text{Information Gain} = 1 - (0.9182 \cdot \frac{3}{4} + 0 \cdot \frac{1}{4}) = 0.3112 \]

The Gain of Information is 0.3112

Table 2 - Table of quantity and proportion for Bumpiness
Steep Flat psteep pflat
Bumpy 1 1 1/2 1/2
Smooth 1 1 1/2 1/2
  • Entropy calculation:

\[entropy(bumpy) = -(p_{bumpy,slow} \cdot \ln(p_{bumpy,slow}) + p_{bumpy,fast} \cdot \ln(p_{bumpy,fast})) = \\ -(0.5 \cdot \ln(0.5) + 0.5 \cdot \ln(0.5)) = 1 \]

\[entropy(smooth) = -(p_{smooth,slow} \cdot \ln(p_{smooth,slow}) + p_{smooth,fast} \cdot \ln(p_{smooth,fast})) = -(0.5 \cdot \ln(0.5) + 0.5 \cdot \ln(0.5)) = 1 \]

  • Information Gain:

\[\text{Information Gain} = 1 - (1 \cdot \frac{2}{4} + 1 \cdot \frac{2}{4}) = 0\]

For this variable there is no Gain of Information, which means we do not learn anything splitting by bumpiness.

Table 3 - Table of quantity and proportion for Speed Limit
Steep Flat psteep pflat
yes 2 0 2/2 0/2
no 0 2 0/2 2/2
  • Entropy Calculation:

\[entropy(yes) = -(p_{yes,slow} \cdot \ln(p_{yes,slow}) + p_{yes,fast} \cdot \ln(p_{yes,fast})) = \\ -(1 \cdot \ln(1) + 0 \cdot \ln(0)) = 0 \]

\[entropy(no) = -(p_{no,slow} \cdot \ln(p_{no,slow}) + p_{no,fast} \cdot \ln(p_{no,fast})) = -(0 \cdot \ln(0) + 1 \cdot \ln(1)) = 0 \]

  • Information Gain:

\[\text{Information Gain} = 1 - (1 \cdot 0 + 1 \cdot 0) = 1\]

Best information gain possible. You should split using this feature.

c5_l4_06.png

Figure 6 - Speed Limit Decision Tree.


5.4.2.5 Bias

Bias in machine learning, means an algorithm that almost “ignores” the data, it has almost no capacity to learn anything.

On the other hand, a machine learning very perceptive to data, will only replicate what it has seen before, it has no capacity to “generalize” a new situation. It is called algorithm with high variance.

What we are looking for is something in the middle of bias and high variance. This is the trade-off of bias-variance.

 

A work by AH Uyekita

anderson.uyekita[at]gmail.com