|
1
|
|
|
""" |
|
2
|
|
|
This example shows how to select the best features for a model |
|
3
|
|
|
and dataset. |
|
4
|
|
|
|
|
5
|
|
|
The boston dataset has 13 features, therefore we have 13 search space |
|
6
|
|
|
dimensions for the feature selection. |
|
7
|
|
|
|
|
8
|
|
|
The function "get_feature_indices" returns the list of features that |
|
9
|
|
|
where selected. This can be used to select the subset of features in "x_new". |
|
10
|
|
|
""" |
|
11
|
|
|
|
|
12
|
|
|
import numpy as np |
|
13
|
|
|
import itertools |
|
14
|
|
|
from sklearn.datasets import load_diabetes |
|
15
|
|
|
from sklearn.model_selection import cross_val_score |
|
16
|
|
|
from sklearn.neighbors import KNeighborsRegressor |
|
17
|
|
|
from hyperactive import Hyperactive |
|
18
|
|
|
from hyperactive.optimizers import EvolutionStrategyOptimizer |
|
19
|
|
|
|
|
20
|
|
|
|
|
21
|
|
|
data = load_diabetes() |
|
22
|
|
|
X, y = data.data, data.target |
|
23
|
|
|
|
|
24
|
|
|
|
|
25
|
|
|
# helper function that returns the selected training data features by index |
|
26
|
|
|
def get_feature_indices(opt): |
|
27
|
|
|
feature_indices = [] |
|
28
|
|
|
for key in opt.keys(): |
|
29
|
|
|
if "feature" not in key: |
|
30
|
|
|
continue |
|
31
|
|
|
if opt[key] == 0: |
|
32
|
|
|
continue |
|
33
|
|
|
|
|
34
|
|
|
nth_feature = int(key.rsplit(".", 1)[1]) |
|
35
|
|
|
feature_indices.append(nth_feature) |
|
36
|
|
|
|
|
37
|
|
|
return feature_indices |
|
38
|
|
|
|
|
39
|
|
|
|
|
40
|
|
|
def model(opt): |
|
41
|
|
|
feature_indices = get_feature_indices(opt) |
|
42
|
|
|
if len(feature_indices) == 0: |
|
43
|
|
|
return 0 |
|
44
|
|
|
|
|
45
|
|
|
feature_idx_list = [idx for idx in feature_indices if idx is not None] |
|
46
|
|
|
x_new = X[:, feature_idx_list] |
|
47
|
|
|
|
|
48
|
|
|
knr = KNeighborsRegressor(n_neighbors=opt["n_neighbors"]) |
|
49
|
|
|
scores = cross_val_score(knr, x_new, y, cv=5) |
|
50
|
|
|
score = scores.mean() |
|
51
|
|
|
|
|
52
|
|
|
return score |
|
53
|
|
|
|
|
54
|
|
|
|
|
55
|
|
|
# each feature is used for training (1) or not used for training (0) |
|
56
|
|
|
search_space = { |
|
57
|
|
|
"n_neighbors": list(range(1, 100)), |
|
58
|
|
|
"feature.0": [1, 0], |
|
59
|
|
|
"feature.1": [1, 0], |
|
60
|
|
|
"feature.2": [1, 0], |
|
61
|
|
|
"feature.3": [1, 0], |
|
62
|
|
|
"feature.4": [1, 0], |
|
63
|
|
|
"feature.5": [1, 0], |
|
64
|
|
|
"feature.6": [1, 0], |
|
65
|
|
|
"feature.7": [1, 0], |
|
66
|
|
|
"feature.8": [1, 0], |
|
67
|
|
|
"feature.9": [1, 0], |
|
68
|
|
|
} |
|
69
|
|
|
|
|
70
|
|
|
|
|
71
|
|
|
optimizer = EvolutionStrategyOptimizer(rand_rest_p=0.20) |
|
72
|
|
|
|
|
73
|
|
|
hyper = Hyperactive() |
|
74
|
|
|
hyper.add_search( |
|
75
|
|
|
model, |
|
76
|
|
|
search_space, |
|
77
|
|
|
n_iter=200, |
|
78
|
|
|
initialize={"random": 15}, |
|
79
|
|
|
optimizer=optimizer, |
|
80
|
|
|
) |
|
81
|
|
|
hyper.run() |
|
82
|
|
|
|
SimonBlanke /
Hyperactive
