Course-work/selector_methods.py at master · rmontanana/Course-work · 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
import numpy as np
from sklearn.neighbors import KDTree
from sklearn.feature_selection import f_classif


class MutualInformation():
    def __init__(self,
                 continuous_variables='two_continuous'):
        self.continuous_variables = continuous_variables

    def mutual_info(self, x, y):
        if (self.continuous_variables == 'two_continuous'):
            hx, bx = histogram(x, bins = int(np.sqrt(x.shape[0])), density=True)
            x_categorical = digitize(x, bx)
            hy, by = histogram(y, bins = int(np.sqrt(y.shape[0])), density=True)
            y_categorical = digitize(y, by)
            return mutual_info_score(x_categorical, y_categorical)
        else:
            return self.entropy(y) - self.conditional_entropy(x, y)

    def entropy(self, y):

        p_y = self.compute_distribution(y)

        self.entropy_ = 0.0
        for k, v in p_y.items():
            self.entropy_ += v * log2(v)
        self.entropy_ = -self.entropy_

        return self.entropy_

    def conditional_entropy(self, x, y):
        hx, bx = histogram(x, bins = int(np.sqrt(x.shape[0])), density=True)
        x_categorical = digitize(x, bx)
        p_x = self.compute_distribution(x_categorical)

        y_categorical = y
        p_y = self.compute_distribution(y)

        self.cond_entropy_ = 0.0
        tmp = 0.0
        for y_ in set(y_categorical):
            x1 = x[y_categorical == y_]
            cond_p_xy = self.compute_distribution(digitize(x1, bx)) #p(x/ y)
            for k, v in cond_p_xy.items():
                self.cond_entropy_ += (v * p_y[y_] * (log2(p_x[k]) - log2(v * p_y[y_])))
        return self.cond_entropy_

    def compute_distribution(self, x):
        d = defaultdict(int)
        for e in x: d[e] += 1
        s = float(sum(d.values()))
        a = dict((k, x / s) for k, x in d.items())
        return a


class Selector():

    def __init__(self, criterion='correlation'):
        self.criterion = criterion

    def fit(self, X, y):

        def mutual_scores(self, X, y):
            scores_ = np.zeros(X.shape[1])
            m = MutualInformation(continuous_variables='one_continuous')

            for i in range(X.shape[1]):
                x = X[:, i]
                scores_[i] = m.mutual_info(x, y)

            return scores_

        def correlation_coef(X, y):
            return np.mean((X - X.mean(axis=0)) * (X - y.mean()), axis=0) \
                            / np.sqrt(X.std(axis=0) * y.std())

        def relief_criterion(X, y):

            num_classes = np.unique(y)

            k = 5

            for c in num_classes:
                x_c = X[y == c]
                k = min(k, x_c.shape[0] - 1)

            same_class = list()
            different_class = list()
            for c in num_classes:
                x_c = X[y == c]
                tree = KDTree(x_c)
                inds_c = np.where(y == c)[0]
                nearest_neighbor = tree.query(x_c, k=k + 1, return_distance=False)[:, 1:]
                same_class.append((inds_c, inds_c[nearest_neighbor]))

                x_not_c = X[y != c]
                inds_c = np.where(y != c)[0]
                tree = KDTree(x_not_c)
                nearest_neighbor = tree.query(x_c, k=k + 1, return_distance=False)[:, 1:]
                different_class.append((inds_c, inds_c[nearest_neighbor]))

            scores_ = np.zeros(X.shape[1])

            for k_ in range(k):
                for c in num_classes:
                    x_c = X[y == c]
                    ind_diff = different_class[c][1][:, k_]
                    ind_same = same_class[c][1][:, k_]
                    scores_ += (np.abs(x_c - X[ind_diff]) / \
                               (np.abs(x_c - X[ind_same]) + 1e-10)).sum(axis=0)
            return scores_

        if (self.criterion == 'correlation'):
            self.scores_ = np.abs(correlation_coef(X, y))

        if (self.criterion == 'mutual_info'):
            self.scores_ = self.mutual_scores(X, y)

        if (self.criterion == 'f_classif'):
            self.scores_ = f_classif(X, y)[0]

        if (self.criterion == 'relief_criterion'):
            self.scores_ = relief_criterion(X, y)

        return self