Multinomial-Random-Forest/Tree.py at main · jiawangbai/Multinomial-Random-Forest · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
# encoding: utf-8

from utils import *
from DecisionNode import *
import numpy as np
from sklearn.model_selection import train_test_split


class Tree:

    def __init__(self, min_samples_leaf=5, partition_rate=1, B1=5, B2=5, B3=None):
        self.root = None
        self.min_samples_leaf = min_samples_leaf
        self.features_attr = None
        self.b_1 = B1/2                            # Factor that influence selection randomness of splitting feature.
        self.b_2 = B2/2                            # Factor that influence selection randomness of splitting point.
        self.b_3 = B3/2 if B3 != None else B3      # Factor that influence labels in leaf nodes.
        self.partition_rate = partition_rate
        self.criterion = cal_gini

        assert self.b_1 >= 0
        assert self.b_2 >= 0
        assert self.b_3 == None or self.b_3 >= 0

    def fit(self, X, y, features_attr=None):

        # feature_atrr: an array, its size is same as the number of features,
        # 'd': discrete, 'c': continuous
        assert len(features_attr) == X.shape[1]

        np.random.seed()

        self.features_attr = features_attr

        X_e, X_s, y_e, y_s = train_test_split(X, y, test_size=self.partition_rate/(self.partition_rate+1))
        structure_points = np.concatenate((np.array(X_s), np.array([y_s]).T), axis=1)
        estimation_points = np.concatenate((np.array(X_e), np.array([y_e]).T), axis=1)

        self.root = self.__build_tree(structure_points, estimation_points)

    def predict(self, X):

        if np.ndim(X) == 1:
            return self.__predict_rec(X, self.root)
        else:
            result = []
            for sample in X:
                result.append(self.__predict_rec(sample, self.root))
            return np.array(result)

    def __predict_rec(self, X, node):
        if node.label is not None:
            return node.label
        else:
            feat_value = X[node.feature]
            feat_attr = self.features_attr[node.feature]
            threshold = node.threshold

            if feat_value is None or feat_value is np.nan:
                choice = np.random.randint(1, 3)
                if choice == 1:
                    return self.__predict_rec(X, node.true_branch)
                else:
                    return self.__predict_rec(X, node.false_branch)
            else:
                if feat_attr == 'd':
                    if feat_value == threshold:
                        return self.__predict_rec(X, node.true_branch)
                    else:
                        return self.__predict_rec(X, node.false_branch)
                elif feat_attr == 'c':
                    if feat_value >= threshold:
                        return self.__predict_rec(X, node.true_branch)
                    else:
                        return self.__predict_rec(X, node.false_branch)

    def __split(self, dataset, split_feature, threshold):

        true_index = []
        false_index = []

        if self.features_attr[split_feature] == 'd':
            for i in range(len(dataset)):
                if dataset[i][split_feature] == threshold:
                    true_index.append(i)
                else:
                    false_index.append(i)
        elif self.features_attr[split_feature] == 'c':
            for i in range(len(dataset)):
                if dataset[i][split_feature] >= threshold:
                    true_index.append(i)
                else:
                    false_index.append(i)

        return true_index, false_index

    def __split_pair(self, dataset, candidate_features):

        current = self.criterion(dataset[:, -1])

        ret = {}

        for feat in candidate_features:
            col = dataset[:, feat]
            unique_col = np.unique(col)
            attr = self.features_attr[feat]
            ret[feat] = []

            threshold_list = []
            if attr == 'd' or unique_col.shape == 1:
                threshold_list = unique_col
            elif attr == 'c':
                threshold_list = [(unique_col[i]+unique_col[i+1]) / 2 for i in range(len(unique_col)-1)]

            for t in threshold_list:
                true_index, false_index = self.__split(dataset, feat, t)

                p = float(len(true_index)) / len(dataset)
                gain = current - p * self.criterion(dataset[true_index, -1]) - \
                       (1-p) * self.criterion(dataset[false_index, -1])

                ret[feat].append([gain, t])
            ret[feat] = np.array(ret[feat])
            ret[feat] = ret[feat][np.argsort(-ret[feat][:, 0])]

        return ret

    def __build_tree(self, structure_points, estimation_points):

        if len(cal_label_dic(structure_points[:, -1])) == 1:
            return DecisionNode(label=voting(cal_label_dic(estimation_points[:, -1])))

        candidate_features = []
        for i in range(structure_points.shape[1]-1):
            if len(np.unique(structure_points[:, i])) > 1:
                candidate_features.append(i)
        if candidate_features == []:
            return DecisionNode(label=voting(cal_label_dic(estimation_points[:, -1])))

        info_gain_dict = self.__split_pair(structure_points, candidate_features)

        info_gain_feat_max = []
        for key, val in info_gain_dict.items():
            info_gain_feat_max.append([key, val[0][0]])
        info_gain_feat_max = np.array(info_gain_feat_max)
        info_gain_feat_max = info_gain_feat_max[np.argsort(-info_gain_feat_max[:, 1])]

        info_gain_feat_max_norm = self.b_1 * max_min_normalization(info_gain_feat_max[:, 1])
        split_feature = int(info_gain_feat_max[mutinomial(info_gain_feat_max_norm)][0])

        info_gain_chosen_feat_norm = self.b_2 * max_min_normalization(info_gain_dict[split_feature][:, 0])
        threshold = info_gain_dict[split_feature][mutinomial(info_gain_chosen_feat_norm)][1]

        true_index_s, false_index_s = self.__split(structure_points, split_feature, threshold)
        true_index_e, false_index_e = self.__split(estimation_points, split_feature, threshold)

        if len(true_index_e) == 0 or len(false_index_e) == 0:
            return DecisionNode(label=voting(cal_label_dic(estimation_points[:, -1]), self.b_3))

        if len(true_index_e) <= self.min_samples_leaf:
            true_branch = DecisionNode(label=voting(cal_label_dic(estimation_points[true_index_e, -1])))
        else:
            true_branch = self.__build_tree(structure_points[true_index_s], estimation_points[true_index_e])

        if len(false_index_e) <= self.min_samples_leaf:
            false_branch = DecisionNode(label=voting(cal_label_dic(estimation_points[false_index_e, -1])))
        else:
            false_branch = self.__build_tree(structure_points[false_index_s], estimation_points[false_index_e])

        return DecisionNode(feature=split_feature, threshold=threshold,
                            true_branch=true_branch, false_branch=false_branch)