|
1
|
|
|
""" |
|
2
|
|
|
This example shows how you can search for the best models in each layer in a |
|
3
|
|
|
stacking ensemble. |
|
4
|
|
|
|
|
5
|
|
|
We want to create a stacking ensemble with 3 layers: |
|
6
|
|
|
- a top layer with one model |
|
7
|
|
|
- a middle layer with multiple models |
|
8
|
|
|
- a bottom layer with multiple models |
|
9
|
|
|
|
|
10
|
|
|
We also want to know how many models should be used in the middle and bottom layer. |
|
11
|
|
|
For that we can use the helper function "get_combinations". It works as follows: |
|
12
|
|
|
|
|
13
|
|
|
input = [1, 2 , 3] |
|
14
|
|
|
output = get_combinations(input, comb_len=2) |
|
15
|
|
|
output: [[1, 2], [1, 3], [2, 3], [1, 2, 3]] |
|
16
|
|
|
|
|
17
|
|
|
Instead of numbers we insert models into "input". This way we get each combination |
|
18
|
|
|
with more than 2 elements. Only 1 model per layer would not make much sense. |
|
19
|
|
|
|
|
20
|
|
|
The ensemble itself is created via the package "mlxtend" in the objective-function "stacking". |
|
21
|
|
|
""" |
|
22
|
|
|
|
|
23
|
|
|
import itertools |
|
24
|
|
|
|
|
25
|
|
|
from sklearn.datasets import load_breast_cancer |
|
26
|
|
|
from sklearn.model_selection import cross_val_score |
|
27
|
|
|
from mlxtend.classifier import StackingClassifier |
|
28
|
|
|
|
|
29
|
|
|
from sklearn.ensemble import ( |
|
30
|
|
|
GradientBoostingClassifier, |
|
31
|
|
|
RandomForestClassifier, |
|
32
|
|
|
ExtraTreesClassifier, |
|
33
|
|
|
) |
|
34
|
|
|
|
|
35
|
|
|
from sklearn.neighbors import KNeighborsClassifier |
|
36
|
|
|
from sklearn.neural_network import MLPClassifier |
|
37
|
|
|
from sklearn.gaussian_process import GaussianProcessClassifier |
|
38
|
|
|
from sklearn.tree import DecisionTreeClassifier |
|
39
|
|
|
from sklearn.naive_bayes import GaussianNB |
|
40
|
|
|
|
|
41
|
|
|
from sklearn.linear_model import LogisticRegression |
|
42
|
|
|
from sklearn.linear_model import RidgeClassifier |
|
43
|
|
|
|
|
44
|
|
|
from hyperactive import Hyperactive |
|
45
|
|
|
|
|
46
|
|
|
data = load_breast_cancer() |
|
47
|
|
|
X, y = data.data, data.target |
|
48
|
|
|
|
|
49
|
|
|
# define models that are used in search space |
|
50
|
|
|
gbc = GradientBoostingClassifier() |
|
51
|
|
|
rfc = RandomForestClassifier() |
|
52
|
|
|
etc = ExtraTreesClassifier() |
|
53
|
|
|
|
|
54
|
|
|
mlp = MLPClassifier() |
|
55
|
|
|
gnb = GaussianNB() |
|
56
|
|
|
gpc = GaussianProcessClassifier() |
|
57
|
|
|
dtc = DecisionTreeClassifier() |
|
58
|
|
|
knn = KNeighborsClassifier() |
|
59
|
|
|
|
|
60
|
|
|
lr = LogisticRegression() |
|
61
|
|
|
rc = RidgeClassifier() |
|
62
|
|
|
|
|
63
|
|
|
|
|
64
|
|
|
def stacking(opt): |
|
65
|
|
|
lvl_1_ = opt["lvl_1"]() |
|
66
|
|
|
lvl_0_ = opt["lvl_0"]() |
|
67
|
|
|
top_ = opt["top"]() |
|
68
|
|
|
|
|
69
|
|
|
stack_lvl_0 = StackingClassifier(classifiers=lvl_0_, meta_classifier=top_) |
|
70
|
|
|
stack_lvl_1 = StackingClassifier(classifiers=lvl_1_, meta_classifier=stack_lvl_0) |
|
71
|
|
|
scores = cross_val_score(stack_lvl_1, X, y, cv=3) |
|
72
|
|
|
|
|
73
|
|
|
return scores.mean() |
|
74
|
|
|
|
|
75
|
|
|
|
|
76
|
|
|
# helper function to create search space dimensions |
|
77
|
|
|
def get_combinations(models, comb_len=2): |
|
78
|
|
|
def _list_in_list_of_lists(list_, list_of_lists): |
|
79
|
|
|
for list__ in list_of_lists: |
|
80
|
|
|
if set(list_) == set(list__): |
|
81
|
|
|
return True |
|
82
|
|
|
|
|
83
|
|
|
comb_list = [] |
|
84
|
|
|
for i in range(0, len(models) + 1): |
|
85
|
|
|
for subset in itertools.permutations(models, i): |
|
86
|
|
|
if len(subset) < comb_len: |
|
87
|
|
|
continue |
|
88
|
|
|
if _list_in_list_of_lists(subset, comb_list): |
|
89
|
|
|
continue |
|
90
|
|
|
|
|
91
|
|
|
comb_list.append(list(subset)) |
|
92
|
|
|
|
|
93
|
|
|
comb_list_f = [] |
|
94
|
|
|
for comb_ in comb_list: |
|
95
|
|
|
|
|
96
|
|
|
def _func_(): |
|
97
|
|
|
return comb_ |
|
|
|
|
|
|
98
|
|
|
|
|
99
|
|
|
_func_.__name__ = str(i) + "___" + str(comb_) |
|
|
|
|
|
|
100
|
|
|
comb_list_f.append(_func_) |
|
101
|
|
|
|
|
102
|
|
|
return comb_list_f |
|
103
|
|
|
|
|
104
|
|
|
|
|
105
|
|
|
def lr_f(): |
|
106
|
|
|
return lr |
|
107
|
|
|
|
|
108
|
|
|
|
|
109
|
|
|
def dtc_f(): |
|
110
|
|
|
return dtc |
|
111
|
|
|
|
|
112
|
|
|
|
|
113
|
|
|
def gnb_f(): |
|
114
|
|
|
return gnb |
|
115
|
|
|
|
|
116
|
|
|
|
|
117
|
|
|
def rc_f(): |
|
118
|
|
|
return rc |
|
119
|
|
|
|
|
120
|
|
|
|
|
121
|
|
|
models_0 = [gpc, dtc, mlp, gnb, knn] |
|
122
|
|
|
models_1 = [gbc, rfc, etc] |
|
123
|
|
|
|
|
124
|
|
|
stack_lvl_0_clfs = get_combinations(models_0) |
|
125
|
|
|
stack_lvl_1_clfs = get_combinations(models_1) |
|
126
|
|
|
|
|
127
|
|
|
|
|
128
|
|
|
print("\n stack_lvl_0_clfs \n", stack_lvl_0_clfs, "\n") |
|
129
|
|
|
|
|
130
|
|
|
|
|
131
|
|
|
search_space = { |
|
132
|
|
|
"lvl_1": stack_lvl_1_clfs, |
|
133
|
|
|
"lvl_0": stack_lvl_0_clfs, |
|
134
|
|
|
"top": [lr_f, dtc_f, gnb_f, rc_f], |
|
135
|
|
|
} |
|
136
|
|
|
|
|
137
|
|
|
""" |
|
138
|
|
|
hyper = Hyperactive() |
|
139
|
|
|
hyper.add_search(stacking, search_space, n_iter=3) |
|
140
|
|
|
hyper.run() |
|
141
|
|
|
""" |
|
142
|
|
|
|
SimonBlanke /
Hyperactive
Loading history...