-
Notifications
You must be signed in to change notification settings - Fork 94
/
matrixfactorization.py
197 lines (155 loc) · 9.35 KB
/
matrixfactorization.py
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
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
"""Collaborative filtering features using various techniques of Matrix Factorization for recommendations.
Recommended for large data"""
"""
Add the user column name and item column name in recipe_dict in config to match the
column names as per the dataset or use the default 'user' and 'item' respectively in your dataset
Sample Datasets
# Netflix - https://www.kaggle.com/netflix-inc/netflix-prize-data
recipe_dict = "{'user_col': 'user', 'item_col': 'movie'}"
# MovieLens - https://grouplens.org/datasets/movielens/
recipe_dict = "{'user_col': 'userId', 'item_col': 'movieId'}"
# RPackages - https://www.kaggle.com/c/R/data
recipe_dict = "{'user_col': 'User', 'item_col': 'Package'}"
"""
import datatable as dt
import numpy as np
import pandas as pd
import h2o4gpu
import scipy
from h2oaicore.systemutils import config
from h2oaicore.transformer_utils import CustomTransformer
from h2oaicore.separators import extra_prefix, orig_feat_prefix, col_sep
from sklearn.decomposition import NMF
from sklearn.model_selection import KFold, StratifiedKFold
from sklearn.preprocessing import LabelEncoder
class RecH2OMFTransformer(CustomTransformer):
_allow_transform_to_modify_output_feature_names = True
_multiclass = False
_unsupervised = False # uses target
_uses_target = True # uses target
_can_use_gpu = True
_mf_type = "h2o4gpu"
def __init__(self, n_components=50, _lambda=0.1, batches=1, max_iter=100, alpha=0.1, **kwargs):
super().__init__(**kwargs)
self.user_col = config.recipe_dict['user_col'] if "user_col" in config.recipe_dict else "user"
self.item_col = config.recipe_dict['item_col'] if "item_col" in config.recipe_dict else "item"
if self.__class__._mf_type == "h2o4gpu":
self._n_components = n_components
self._lambda = _lambda
self._batches = batches
self._max_iter = max_iter
elif self.__class__._mf_type == "nmf":
self._n_components = n_components
self._alpha = alpha
self._max_iter = max_iter
@staticmethod
def do_acceptance_test():
return False
@staticmethod
def get_default_properties():
return dict(col_type="all", min_cols="all", max_cols="all", relative_importance=1, num_default_instances=1)
@staticmethod
def get_parameter_choices():
return {"n_components": [10, 30, 50, 70, 100],
"_lambda": [0.01, 0.05, 0.1],
"batches": [1],
"max_iter": [10, 50, 100, 200],
"alpha": [0.01, 0.05, 0.1]}
def fit_transform(self, X: dt.Frame, y: np.array = None):
if len(np.unique(self.labels)) == 2:
le = LabelEncoder()
self.labels = le.fit_transform(self.labels)
y = np.array(le.transform(y), dtype="float32")
else:
y = np.array(y, dtype="float32")
X = X[:, [self.user_col, self.item_col]]
self.user_le = LabelEncoder()
self.item_le = LabelEncoder()
X[:, self.user_col] = dt.Frame(self.user_le.fit_transform(X[:, self.user_col]))
X[:, self.item_col] = dt.Frame(self.item_le.fit_transform(X[:, self.item_col]))
X_pd = X.to_pandas()
if len(np.unique(self.labels)) == 2:
kfold = StratifiedKFold(n_splits=10)
else:
kfold = KFold(n_splits=10)
preds = np.full(X.nrows, fill_value=np.nan)
for train_index, val_index in kfold.split(X_pd, y):
X_train, y_train = X_pd.iloc[train_index,], y[train_index]
X_val, y_val = X_pd.iloc[val_index,], y[val_index]
X_val2 = X_val[(X_val[self.user_col].isin(np.unique(X_train[self.user_col]))) & (
X_val[self.item_col].isin(np.unique(X_train[self.item_col])))]
y_val2 = y_val[(X_val[self.user_col].isin(np.unique(X_train[self.user_col]))) & (
X_val[self.item_col].isin(np.unique(X_train[self.item_col])))]
X_panel = pd.concat([X_train, X_val2], axis=0)
users, user_indices = np.unique(np.array(X_panel[self.user_col], dtype="int32"), return_inverse=True)
items, item_indices = np.unique(np.array(X_panel[self.item_col], dtype="int32"), return_inverse=True)
X_train_user_item_matrix = scipy.sparse.coo_matrix(
(y_train, (user_indices[:len(X_train)], item_indices[:len(X_train)])), shape=(len(users), len(items)))
X_train_shape = X_train_user_item_matrix.shape
X_val_user_item_matrix = scipy.sparse.coo_matrix(
(np.ones(len(X_val2), dtype="float32"), (user_indices[len(X_train):], item_indices[len(X_train):])),
shape=X_train_shape)
if self.__class__._mf_type == "h2o4gpu":
factorization = h2o4gpu.solvers.FactorizationH2O(self._n_components, self._lambda,
max_iter=self._max_iter)
factorization.fit(X_train_user_item_matrix, X_BATCHES=self._batches, THETA_BATCHES=self._batches)
preds[val_index[(X_val[self.user_col].isin(np.unique(X_train[self.user_col]))) & (
X_val[self.item_col].isin(np.unique(X_train[self.item_col])))]] = factorization.predict(
X_val_user_item_matrix).data
elif self.__class__._mf_type == "nmf":
factorization = NMF(n_components=self._n_components, alpha=self._alpha, max_iter=self._max_iter)
user_matrix = factorization.fit_transform(X_train_user_item_matrix)
item_matrix = factorization.components_.T
val_users = np.take(user_matrix, X_val_user_item_matrix.row, axis=0)
val_items = np.take(item_matrix, X_val_user_item_matrix.col, axis=0)
preds[val_index[(X_val[self.user_col].isin(np.unique(X_train[self.user_col]))) & (
X_val[self.item_col].isin(np.unique(X_train[self.item_col])))]] = np.sum(val_users * val_items,
axis=1)
users, user_indices = np.unique(np.array(X_pd[self.user_col], dtype="int32"), return_inverse=True)
items, item_indices = np.unique(np.array(X_pd[self.item_col], dtype="int32"), return_inverse=True)
X_train_user_item_matrix = scipy.sparse.coo_matrix((y, (user_indices, item_indices)), shape=(len(users), len(items)))
self.X_train_shape = X_train_user_item_matrix.shape
if self.__class__._mf_type == "h2o4gpu":
self.factorization = h2o4gpu.solvers.FactorizationH2O(self._n_components, self._lambda,
max_iter=self._max_iter)
self.factorization.fit(X_train_user_item_matrix, X_BATCHES=self._batches, THETA_BATCHES=self._batches)
elif self.__class__._mf_type == "nmf":
factorization = NMF(n_components=self._n_components, alpha=self._alpha, max_iter=self._max_iter)
self.user_matrix = factorization.fit_transform(X_train_user_item_matrix)
self.item_matrix = factorization.components_.T
# output feature names
if self.__class__._mf_type == "h2o4gpu":
self._output_feature_names = [(f"{self.display_name}{orig_feat_prefix}{self.user_col}{col_sep}"
f"{self.item_col}.n_components={self._n_components},"
f"lambda={self._lambda},batches={self._batches},max_iter={self._max_iter}")]
elif self.__class__._mf_type == "nmf":
self._output_feature_names = [(f"{self.display_name}{orig_feat_prefix}{self.user_col}{col_sep}"
f"{self.item_col}.n_components={self._n_components},"
f"alpha={self._alpha},max_iter={self._max_iter}")]
# output feature descriptions
self._feature_desc = [f"Recommender transformer ({self.__class__._mf_type}): " + self._output_feature_names[0]]
return preds
def transform(self, X: dt.Frame):
X = X[:, [self.user_col, self.item_col]]
preds = np.full(X.nrows, fill_value=np.nan)
X_pd = X.to_pandas()
X_test = X_pd[
(X_pd[self.user_col].isin(self.user_le.classes_)) & (X_pd[self.item_col].isin(self.item_le.classes_))]
X_test[self.user_col] = self.user_le.transform(X_test[self.user_col])
X_test[self.item_col] = self.item_le.transform(X_test[self.item_col])
X_test_user_item_matrix = scipy.sparse.coo_matrix(
(np.ones(len(X_test), dtype="float32"), (X_test[self.user_col], X_test[self.item_col])),
shape=self.X_train_shape)
if self.__class__._mf_type == "h2o4gpu":
preds[(X_pd[self.user_col].isin(self.user_le.classes_)) & (
X_pd[self.item_col].isin(self.item_le.classes_))] = self.factorization.predict(
X_test_user_item_matrix).data
elif self.__class__._mf_type == "nmf":
test_users = np.take(self.user_matrix, X_test_user_item_matrix.row, axis=0)
test_items = np.take(self.item_matrix, X_test_user_item_matrix.col, axis=0)
preds[(X_pd[self.user_col].isin(self.user_le.classes_)) & (
X_pd[self.item_col].isin(self.item_le.classes_))] = np.sum(test_users * test_items, axis=1)
return preds
class RecNMFTransformer(RecH2OMFTransformer):
_can_use_gpu = False
_mf_type = "nmf"