Step-by-Step Guide To Implement Machine Learning II - Decision Tree

Ryukkkk

4.00/5 (4 votes)

May 5, 2019

CPOL

3 min read

7640

Easy to implement machine learning

Introduction

The principle of decision is not complex. Start from the root node, compare the feature value stored in node with the corresponding feature value of the test object. Then, turn to left substree or right substree recursively according to the comparison result. Finally, the label of the leaf node is the prediction of the test object.

For example, there is a decision tree below, where the feature set is {hungry, money} and the label set is {go to restaurant, go to buy a hamburger, go to sleep}.

In the decision process, first if I am hungry, turn to the right subtree, which means I will go to sleep. Else turn to the left subtree. Then, check how much money in my wallet. If I have more than 25$, I will go to restaurant. Else, I can only go to buy a hamburger.

Decision Tree Model

Decision Tree model consists of feature selection, generation of decision tree and classify.

Feature Selection

The aim of feature selection is to choose the feature which has the ability of classification for training data to make the learning process more efficient. Feature selection includes two parts, namely, feature selection and feature value selection. The selected tuple (feature, feature value) are applied as the node in the decision tree.

Feature selection is usually based on information gain or information gain ratio. Information gain is defined as the difference between the empirical entropy of set D, $H(D)$ and the conditional entropy under selecting feature A, $H\left( D|A \right)$ , which is calculated as:

$g\left( D,A \right)=H\left( D \right)-H\left( D|A \right)$

Specifically, when the data set D is divided into two subset by feature A, the information gain is expressed as:

$g\left( D|A \right)=H\left( D \right)-\alpha H\left( D_{1} \right)-\left( 1-\alpha\right) H\left( D_{2} \right)$

where $\alpha$ is the ratio of the first subset, namely:

$\alpha=\frac{\left|D_{1}\right|}{\left|D\right|}$

The code of calculation of information gain is shown below:

 left_set, right_set = self.divideData(data, i, value)
 # calculate information gain
 ratio = float(len(left_set)/sample_num)
 if ratio == 0.0:
     info_gain = init_entropy - (1 - ratio) * self.getEntropy(right_set[:, -1])
 elif ratio == 1.0:
     info_gain = init_entropy - ratio*self.getEntropy(left_set[:, -1])
 else:
     info_gain = init_entropy - ratio * 
     self.getEntropy(left_set[:, -1]) - (1 - ratio) * self.getEntropy(right_set[:, -1])

So far, we have learned how to calculate the information gain. But, how to calculate the empirical entropy? The empirical entropy is the entropy of the labels of the training set, which is given by:

$H\left( D\right)=-\sum_{k=1}^{K}\frac{\left| C_{k}\right|}{\left| D\right|}log_{2}\frac{\left| C_{k}\right|}{\left|D\right|}$

The above equation looks a bit complex but it is very easy to implement. Let's look at the code of it:

    def getEntropy(self, labels):
        labels_num = len(labels)
        label_count = self.uniqueCount(labels)

        entropy = 0.0
        for j in label_count:
            prop = label_count[j]/labels_num
            entropy = entropy + (-prop*math.log(prop, 2))

        return entropy

Generation of Decision Tree

There are many algorithms to generate decision tree, such as ID3, C4.5. In this article, we take ID3 algorithm as the example to generate the decision tree.

First, let's figure out the split process after feature selection. It is known to us that feature selection is to make the data classifiable. Thus, the split process is to divide the training data according to the selected feature index and its selected value value. Specifically, the split process is that if the index feature value in a sample is larger than value, then add the sample into right subtree and delete the index feature in the sample; else if the index feature value in a sample is smaller than value, then add the sample into left subtree and delete the index feature in the sample. The code of split process is:

    def divideData(self, data, index, value):
        left_set = []
        right_set = []
        # select feature in index with value
        for temp in data:
            if temp[index] >= value:
                # delete this feature
                new_feature = np.delete(temp, index)
                right_set.append(new_feature)
            else:
                new_feature = np.delete(temp, index)
                left_set.append(new_feature)
        return np.array(left_set), np.array(right_set)

Before generating a decision tree, we define a data structure to save the node in the decision tree:

class DecisionNode:
    def __init__(self, index=-1, value=None, results=None, right_tree=None, left_tree=None):
        self.index = index                    # the index of feature
        self.value = value                    # the value of the feature with index
        self.results = results                # current decision result
        self.right_tree = right_tree
        self.left_tree = left_tree

Then, we can generate the decision tree recursively. If there is no feature in the data set, stop. If the information gain is smaller than a given threshold, stop. Else, split the data set according to the best selected feature and its value as shown below:

     def createDecisionTree(self, data):
        # if there is no feature in data, stop division
        if len(data) == 0:
            self.tree_node = DecisionNode()
            return self.tree_node

        best_gain = 0.0
        best_criteria = None
        best_set = None

        feature_num = len(data[0]) - 1
        sample_num = len(data[:, -1])
        init_entropy = self.getEntropy(data[:, -1])

        # get the best division
        for i in range(feature_num):
            uniques = np.unique(data[:, i])
            for value in uniques:
                left_set, right_set = self.divideData(data, i, value)
                # calculate information gain
                ratio = float(len(left_set)/sample_num)
                if ratio == 0.0:
                    info_gain = init_entropy - (1 - ratio) * self.getEntropy(right_set[:, -1])
                elif ratio == 1.0:
                    info_gain = init_entropy - ratio*self.getEntropy(left_set[:, -1])
                else:
                    info_gain = init_entropy - ratio * self.getEntropy
                      (left_set[:, -1]) - (1 - ratio) * self.getEntropy(right_set[:, -1])
                if info_gain > best_gain:
                    best_gain = info_gain
                    best_criteria = (i, value)
                    best_set = (left_set, right_set)

        # create the decision tree
        if best_gain < self.t:
            self.tree_node = DecisionNode(results=self.uniqueCount(data[:, -1]))
            return self.tree_node
        else:
            ltree = self.createDecisionTree(best_set[0])
            rtree = self.createDecisionTree(best_set[1])
            self.tree_node = DecisionNode(index=best_criteria[0], 
                             value=best_criteria[1], left_tree=ltree, right_tree=rtree)
            return self.tree_node

Classify

The principle of classification is like the binary sort tree, namely, comparing the feature value stored in node with the corresponding feature value of the test object. Then, turn to left substree or right substree recursively as shown below:

    def classify(self, sample, tree):
        if tree.results != None:
            return tree.results
        else:
            value = sample[tree.index]
            branch = None
            if value >= tree.value:
                branch = tree.right_tree
            else:
                branch = tree.left_tree
            return self.classify(sample, branch)

Conclusion and Analysis

There exist pruning process by dynamic programming after generation of decision tree to get the best decision tree. Moreover, Classification and Regression Tree (CART) is a type of decision tree which can be not only applied in classification but also regression. Finally, let's compare our Decision tree with the tree in Sklearn and the detection performance is displayed below:

From the figure, we can learn that the decision tree is not as good as sklearn's, which may be that we don't apply the pruning process.

The related code and dataset in this article can be found in MachineLearning.