-
Notifications
You must be signed in to change notification settings - Fork 0
/
msePipeline.py
765 lines (655 loc) · 30.5 KB
/
msePipeline.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
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
"""
author: Colin Bradley
last updated: 05/12/2021
FIXME: Val step in SGD not working.
TODO
----
1. Still improving hyperparam optimization.
2. Parallelize SGD.
"""
import os
import pandas as pd
import numpy as np
from sqlalchemy import create_engine
from scipy import sparse
#########################################################################
############################### CLASSES #################################
#########################################################################
class MSEPipeline():
'''
This class is designed to pull book/user data and perform all the preprocessing necessary to trian
a model. It will pull data from an RDS server, clean up column names and some labels, and split
up test, train and validation sets.
TODO:
1) Make some functions in here private methods.
'''
def __init__(self, deploy=False, ratingsThresh=0):
self.archived_ratings = pd.DataFrame(dtype=np.int8)
self.user_ratings = pd.DataFrame(dtype=np.int8)
self.user_archive_ids = {}
self.deploy = deploy
self.ratingsThresh = ratingsThresh
def preprocess(self):
'''
This function is just a clean way to call all preprocessing steps. These steps
include reading in the goodbooks-10k data, fixing some book id's and more.
'''
self.read_data()
self.remove_ratings_below_thresh()
if self.deploy: # only add site users to df if we are deploying
self.remove_new_users()
self.add_users_to_archive()
self.fix_ids()
def read_data(self):
'''
Reads the data from an Amazon RDS database or the local database depending on DEPLOY value.
TODO
----
1. Add functionality to pull from local database.
- FIXED 03/02/2021:17:35PST
2. Add user functionality in local database.
3. Add ability to pull from local CSV files.
'''
# the following lines set up SQL alchemy to grab data from my RDS database. You may need to adjust this to fit yours.
if self.deploy: # if we are running from RDS server
RDS_HOSTNAME = os.environ.get("RDS_HOSTNAME")
RDS_PORT = os.environ.get("RDS_PORT")
RDS_DB_NAME = os.environ.get("RDS_DB_NAME")
RDS_USERNAME = os.environ.get("RDS_USERNAME")
RDS_PASSWORD = os.environ.get("RDS_PASSWORD")
engine = create_engine(
f"postgresql://{RDS_USERNAME}:{RDS_PASSWORD}@{RDS_HOSTNAME}:{RDS_PORT}/{RDS_DB_NAME}")
# this reason for this chunksize structure is to not use too much RAM.
with engine.connect() as connection:
temp1 = pd.read_sql_table(
'archive_rating', connection, chunksize=10000)
temp2 = pd.read_sql_table(
'user_rating', connection, chunksize=10000)
# now just append to dataframes, uncomment the mem lines if you want to see
# how much memory the dataframes are taking.
for chunk in temp1:
self.archived_ratings = self.archived_ratings.append(chunk)
# amem = self.archived_ratings.memory_usage().sum()/(10**6)
for chunk in temp2:
self.user_ratings = self.user_ratings.append(chunk)
# umem = self.user_ratings.memory_usage().sum()/(10**6)
else: # if we are running locally
DB_USER = os.environ.get("DB_USER")
DB_PASS = os.environ.get("DB_PASS")
DB_NAME = os.environ.get("DB_NAME")
engine = create_engine(
f"postgresql://{DB_USER}:{DB_PASS}@localhost/{DB_NAME}")
# I have plenty of RAM locally, no need for chunks.
with engine.connect() as connection:
self.archived_ratings = pd.read_sql_table(
'archive_rating', connection)
self.user_ratings = pd.read_sql_table(
'user_rating', connection)
def remove_ratings_below_thresh(self):
'''
Here we remove all the archived ratings that are below the given
threshhold.
'''
if self.ratingsThresh != 0:
# just keep archived ratings above the passed threshhold value
self.archived_ratings = self.archived_ratings.query(
'rating >= @self.ratingsThresh')
# reserialize the archive ID's so they are contiguous
temp_uids = reserialize_users(
self.archived_ratings.user_id.unique().tolist())
self.archived_ratings.user_id = self.archived_ratings.user_id.map(
temp_uids)
# note we don't actually care about the original archive ID's so we can just
# leave them in this function.
def remove_new_users(self):
'''
Here we dump the users with less than five ratings. The algorithm works
better the more ratings a user has made.
'''
# check if users have more than 5 reviews
grouped_users = self.user_ratings.groupby('site_id').count()
# delete all users who don't
for idx, group in grouped_users.iterrows():
if group.rating < 6:
self.user_ratings = self.user_ratings[self.user_ratings['site_id'] != idx]
def add_users_to_archive(self):
'''
Here we take our application users reviews (after they've been filtered for low-review users) and tack them onto our existing table of archived ratings.
'''
# this first bit just adds an archive user ID to our users so
# they can be tacked onto the end of the archive user ratings.
last_archived_user = self.archived_ratings.user_id.max()
ulist = self.user_ratings.site_id.unique().tolist()
for i in range(len(ulist)):
arcid = last_archived_user + i + 1
temp = {ulist[i]: arcid}
self.user_archive_ids.update(temp)
self.user_ratings['user_id'] = self.user_ratings.site_id.map(
self.user_archive_ids)
# combine the ratings sets, change column names, fix typing
self.archived_ratings = self.archived_ratings.append(
self.user_ratings[['user_id', 'book_id', 'rating']])
# now we need to create a dataframe that will have every book for each user.
# We will use this in the prediction step, since our predict() method only
# works on book_id's that exist in the dataframe.
bookList = self.archived_ratings.book_id.unique().tolist()
userList = list(self.user_archive_ids.values())
dfList = []
for user in userList:
for book in bookList:
dfList.append([user, book, np.nan])
self.user_predictions = pd.DataFrame(
dfList, columns=['user_id', 'book_id', 'prediction'])
def commit_recommendations(self):
'''
Here we commit the recommendations to the database to be read on the app for users
to enjoy. For this to work you MUST MAKE PREDICTIONS
'''
try:
# just remap ID's to their original format
self.user_predictions.iid = self.user_predictions.iid + 1
self.user_predictions.uid = self.user_predictions.uid + 1
recs = self.user_predictions[self.user_predictions.uid.isin(
list(self.user_archive_ids.values()))]
mapping = {x: y for y, x in self.user_archive_ids.items()}
RDS_HOSTNAME = os.environ.get("RDS_HOSTNAME")
RDS_PORT = os.environ.get("RDS_PORT")
RDS_DB_NAME = os.environ.get("RDS_DB_NAME")
RDS_USERNAME = os.environ.get("RDS_USERNAME")
RDS_PASSWORD = os.environ.get("RDS_PASSWORD")
engine = create_engine(
f"postgresql://{RDS_USERNAME}:{RDS_PASSWORD}@{RDS_HOSTNAME}:{RDS_PORT}/{RDS_DB_NAME}")
with engine.connect() as connection:
connection.execute('DELETE FROM user_recs')
for rec in recs.iterrows():
iid = int(rec[1].iid)
sid = mapping[rec[1].uid]
score = rec[1].prediction
user_books = self.user_ratings[self.user_ratings.site_id == sid].book_id.tolist(
)
if iid not in user_books:
connection.execute(
f"INSERT INTO user_recs(book_id, site_id, score) VALUES ({iid}, {sid}, {score})")
except:
# FIXME This is bad practice. You should handle specific exceptions.
print(
"Something went wrong, maybe you called this method without providing predictions.")
def fix_ids(self):
'''
This function sets the bookand user id's to start at zero. It also changes the name of headers from user_id and book_id to uid and iid.
TODO
----
1. Include the book/user metadata and clean that up as well.
'''
# just changing to standard feature names
self.archived_ratings.rename(
columns={'user_id': 'uid', 'book_id': 'iid'}, inplace=True)
# starting user and book indices from 0
self.archived_ratings.uid = self.archived_ratings.uid - 1
self.archived_ratings.iid = self.archived_ratings.iid - 1
# do the same for user_predictions if we are deploying.
if self.deploy:
self.user_predictions.rename(
columns={'user_id': 'uid', 'book_id': 'iid'}, inplace=True)
self.user_predictions.uid = self.user_predictions.uid - 1
self.user_predictions.iid = self.user_predictions.iid - 1
def split_test_train(self,
testTrainFrac=0.5,
ratingsWithheldFrac=0.4,
testValFrac=0.5):
'''
This function takes in the entire ratings dataset and splits the interaction matrices into a training, testing and validation set.
First the function splits off a portion of reviews that are then split up into test and validation sets. This function is different
from scikit-learn in that it preserves all user-ids in the training set to avoid a cold start problem when testing your algorithm.
Parameters
----------
testTrainFrac : float, optional
What percentage of the users would you like to test/validate on. The default is 0.5.
ratingsWithheldFrac : float, optional
What percentage of the ratings of the test/val users would you like to withhold. The default is 0.4.
testValFrac : float, optional
Fraction to split test into test/validation. The default is 0.5.
Returns
-------
sparseTrain : scipy coo matrix
A sparse matrix of the training user-item interactions.
sparseTest : scipy coo matrix
A sparse matrix of the testing user-item interactions.
sparseVal : scipy coo matrix
A sparse matrix of the validation user-item interactions.
TODO
----
1. Create cross validation functionality where we can split the data into 3-5 sets.
'''
import random
random.seed(1001)
# let's get the list of users for the test set
uids = self.archived_ratings.uid.unique().tolist()
test_uids = random.sample(uids, k=int(len(uids)*testTrainFrac))
# get all data for test users and put the rest in training
test_users = self.archived_ratings.query('uid == @test_uids')
train = self.archived_ratings.query('uid != @test_uids')
# only consider 40% of test users ratings and put the rest in the train set.
# This prevents the cold start problem on the test set. We will later incorporate
# new users.
#
# This just samples ratingsWithheldFrac of the books
test = test_users.groupby('uid').sample(
frac=ratingsWithheldFrac, random_state=1)
# Now let's sample some of these for a validation set, this could be better but I want to get this working for now
validation = test.groupby('uid').sample(
frac=testValFrac, random_state=888)
# lets add the non train/val books to the test set.
train = train.append(test_users.drop(test.index), ignore_index=True)
# now just dropping val set from train
test.drop(validation.index, inplace=True)
trainNumBooks = len(train.iid.unique())
testNumBooks = len(test.iid.unique())
valNumBooks = len(validation.iid.unique())
trainNumUsers = len(train.uid.unique())
testNumUsers = len(test.uid.unique())
valNumUsers = len(validation.uid.unique())
print(
f"The number of books in the train set: {trainNumBooks}, test set: {testNumBooks}, val set: {valNumBooks}. The number of users in the train set: {trainNumUsers}, test set: {testNumUsers}, val set: {valNumUsers}.")
return train, test, validation
class MSErec():
'''
This class will perform all ML processes to predict books for users of our app. It will create
user/item matrices, perform gradient descent (with momentum), and output predictions!
'''
def __init__(self, df, test=None, validation=None):
'''
Parameters
----------
df : pandas DataFrame
Returns
-------
'''
# let's create a class dataframe object first
self.df = df
# FIXME: if you take out users (like splitting test and training) then the user id's and the length of the
num_uid = len(df.uid.unique())
num_iid = len(df.iid.unique())
# create sparse matrices
self.utility = create_sparse_matrix(df, num_uid, num_iid)
# only create matrices for test and val if passed
if test is not None:
self.test = create_sparse_matrix(test, num_uid, num_iid)
else:
self.test = None
if validation is not None:
self.validation = create_sparse_matrix(
validation, num_uid, num_iid)
else:
self.validation = None
def trainModel(self, K=25, epochs=155, gamma=20, lr=0.05):
'''
Parameters
----------
Returns
-------
TODO
----
NOTE: optimal K=25, epochs=155, gamma=20, lr=0.05. Still optimizing but this is ok for initial deployment
NOTE: I've removed momentum for now. Sloppy but I don't need it right now.
'''
# this initializes some embedding matrices
num_uid = self.utility.shape[0]
num_iid = self.utility.shape[1]
self.user_features = create_embeddings(num_uid, K=K, gamma=gamma)
self.item_features = create_embeddings(num_iid, K=K, gamma=gamma)
# now perform GD, check if we passed a validation set as well.
if self.validation is not None:
self.emb_user, self.emb_item, cost_train, cost_val = gradient_descent(df=self.df,
utility=self.utility,
user_features=self.user_features,
item_features=self.item_features,
epochs=epochs,
val=self.validation,
updates=False)
return (cost_train, cost_val)
else:
self.emb_user, self.emb_item, cost_train = gradient_descent(df=self.df,
utility=self.utility,
user_features=self.user_features,
item_features=self.item_features,
epochs=epochs,
updates=False)
return (cost_train,)
def paramSearch(self, num_samples=5):
'''
Parameters
----------
Returns
-------
TODO
----
'''
hyperparams = pd.DataFrame()
print('Searching for optimal parameters...')
for i in range(num_samples):
# get a random sample of hyperparameters
params = sample_hyperparameters()
cost = self.trainModel(K=params["K"],
beta=params["beta"],
epochs=params["epochs"],
gamma=params["gamma"],
lr=params["lr"])
params['train_mse'] = cost[0]
# FIXME: This also won't work because of
# val step in trainModel(), len(cost) will
# never be 2
if len(cost) == 2:
params['val_mse'] = cost[1]
hyperparams = hyperparams.append(params, ignore_index=True)
return hyperparams.sort_values(by='train_mse')
def gridSearch(self, dfParams):
Ks = [20, 25, 30]
epochs = [100, 125, 150]
gammas = [15, 20, 25]
lrs = [0.025, 0.05, 0.075]
for k in Ks:
for gamma in gammas:
for epoch in epochs:
for lr in lrs:
dfError = pd.DataFrame()
# this initializes some embedding matrices
num_uid = self.utility.shape[0]
num_iid = self.utility.shape[1]
self.user_features = create_embeddings(
num_uid, K=k, gamma=gamma)
self.item_features = create_embeddings(
num_iid, K=k, gamma=gamma)
self.emb_user, self.emb_item, cost_train, dfError = gradient_descent(df=self.df,
utility=self.utility,
user_features=self.user_features,
item_features=self.item_features,
epochs=epoch,
learning_rate=lr,
updates=False,
dfError=dfError)
dfParams = dfParams.append(
[[k, gamma, epoch, lr, cost_train]])
dfError.to_csv(
f"AnalyzedData/error_E{epoch}_L{lr}_K{k}_G{gamma}-{pd.to_datetime('today').strftime('%m-%d-%Y')}.csv")
return dfParams
def getPredictions(self, df, num_predict=10):
'''
This is actually more complex than this. By just passing self.df, you're only getting predictions on items already read.
What you need to do is get this to predict on each user. To achieve this, create a dataframe with all books for each new user!
Parameters
----------
Returns
-------
TODO
----
1. Fix this so it works for local DB's as well.
- NOTE: I'm not actually sure this is where the problem is. Probably more in commit_recommendations() under msePipeline().
'''
totalPredictions = predict(df=df,
user_features=self.user_features,
item_features=self.item_features)
grouped = totalPredictions.groupby('uid')
truncatedPredictions = pd.DataFrame()
check = pd.DataFrame()
for group in grouped:
temp = group[1].sort_values(by='prediction',
ascending=False)
truncatedPredictions = truncatedPredictions.append(
temp.iloc[:num_predict])
check = check.append(
temp.iloc[-num_predict:])
print("standard: ", truncatedPredictions,
"reverse: ", check, sep='\n'*2)
return truncatedPredictions
#########################################################################
############################## FUNCTIONS ################################
#########################################################################
def reserialize_users(uids):
'''
If you only consider ratings above a certain number you then lose some users.
Creating a sparse matrix only works with contiguous user id's, so we have to reserialize
the uid's. This function does that.
Parameters
----------
uids: list
A list of the non-contiguous uids
Returns
-------
mapping: dictionary
a mapping between the newly contiguous list of uids and the old list
'''
mapping = {uids[i-1]: i for i in range(1, len(uids)+1)}
return mapping
def create_sparse_matrix(df, rows, cols, column_name="rating"):
'''
Creates a scipy sparse matrix
Parameters
----------
df : pandas DataFrame
The data that will be made a sparse matrix
rows : int
number of rows in the matrix
columns : int
number of columns in the matrix
column_name : str, optional
name of ratings column. This should always be "rating"
Returns
-------
___ : scipy sparse matrix
TODO
----
'''
return sparse.csc_matrix((df[column_name].values, (df['uid'].values, df['iid'].values)), shape=(rows, cols))
def create_embeddings(n, K, gamma=7):
'''
Initializes n x K feature matrices
Parameters
----------
n : int
number of items/users
K : int
number of latent factors, to be tuned
gamma : float, optional
scaling of intial feature weights, to be tuned
Returns
-------
___ : numpy matrix
a randomly initialized feature matrix
TODO
----
'''
return (gamma*np.random.rand(n, K) / K).astype('float32')
def predict(df, user_features, item_features):
'''
This function performs the element wise prediction of each item for each user. It avoids building the
approximated utility matrix in order to save space
Parameters
----------
df : pandas DataFrame
This is the pandas dataframe of the data predictions are to be made on.
user_features : numpy array
The user feature embeddings.
item_features : numpy Array
The item feature embeddings.
Returns
-------
df : pandas DataFrame
The same dataframe as inputted but with a new/updated predictions column.
'''
df['prediction'] = np.sum(np.multiply(
item_features[df['iid']], user_features[df['uid']], dtype=np.float32), axis=1)
return df
def meanSquareError(df, user_features, item_features):
'''
Computes the mean square error on the predictions.
Parameters
----------
df : pandas DataFrame
This is the pandas dataframe of the data predictions are to be made on.
user_features : numpy array
The user feature embeddings.
item_features : numpy Array
The item feature embeddings.
Returns
-------
mse : float
The mean square error for the given embedding matrices.
TODO:
1. Fix this to generalize for any number of features???
'''
# we need to actually make predictions then convert those into a sparse matrix
utility = create_sparse_matrix(
df, user_features.shape[0], item_features.shape[0])
temp = predict(df=df, user_features=user_features,
item_features=item_features)
prediction = create_sparse_matrix(
temp, user_features.shape[0], item_features.shape[0], 'prediction')
# now let's get an error matrix then return the MSE.
error = utility-prediction
mse = (1/len(df))*np.sum(error.power(2))
return mse
def gradient_reg(df, utility, user_features, item_features, lmbda_a, lmbda_b):
'''
Computes the regularized gradient of the mean square error. Returns the gradient
in the 'directions' of both embedded matrices.
Parameters
----------
df : pandas DataFrame
This is the pandas dataframe of the data predictions are to be made on.
utility : scipy sparse matrix
The sparse utility matrix of all of the ratings.
user_features : numpy array
The user feature embeddings.
item_features : numpy Array
The item feature embeddings.
lmbda_a, lmbda_b : float
These parameters are the regularization coefficients.
Returns
-------
grad_user : numpy array
gradient of the MSE, partial derivative w.r.t. the user
grad_item : numpy array
gradient of the MSE, partial derivative w.r.t. the item
'''
# we need to actually make predictions then convert those into a sparse matrix
temp = predict(df=df, user_features=user_features,
item_features=item_features)
prediction = sparse.csc_matrix((temp.prediction.values, (temp.uid.values, temp.iid.values)),
shape=(user_features.shape[0], item_features.shape[0]))
# now let's get an error matrix
error = utility-prediction
# we can now compute the gradient
# we will compute each 'direction' separately and return them separately
grad_user = (-2/df.shape[0]) * \
(error*item_features) + 2*lmbda_a*user_features
grad_item = (-2/df.shape[0])*((error.T) *
user_features) + 2*lmbda_b*item_features
return grad_user, grad_item
def gradient_descent(df,
utility,
user_features,
item_features,
val=None,
lmbda_a=0.002,
lmbda_b=0.002,
epochs=200,
learning_rate=0.05,
beta=0.9,
updates=True,
dfError=None
):
'''
Performs gradient descent to find the optimal embedded matrices. A momentum term
is added to arrive at the minimum sooner. This function will iterate a number of times
specified by the user. It will update the user every 50 epochs on how the cost function
looks. Finally it will return the new embedded matrices and the final cost values.
Parameters
----------
df : pandas DataFrame
This is the pandas dataframe of the data predictions are to be made on.
utility : scipy sparse matrix
The sparse utility matrix of all of the ratings.
user_features : numpy array
The user feature embeddings.
item_features : numpy Array
The item feature embeddings.
val : pandas DataFrame DEFAULT=None
The validation set to check the algorithm against.
lmbda_a, lmbda_b : float, DEFAULT=0.002 for both
These parameters are the regularization coefficients.
epochs : int, DEFAULT=200
The number of iterations on which to perform GD
learning_rate : float, DEFAULT=0.05
The learning rate for GD.
beta : float, DEFAULT=0.9
The momentum coefficient.
updates: bool, DEFAULT=True
The option to print periodic updates of the MSE as the algorithm runs.
Updates will print every epoch with the MSE of the set. It will give
the MSE of the validation set if provided.
Returns
-------
user_features : numpy array
The optimized user feature embeddings.
item_features : numpy Array
The optimized item feature embeddings.
mse_train : float
The final MSE of the training set
mse_val : float, OPTIONAL
the final MSE of the validation set
TODO:
1. Need to re-add momentum to see if that helps with MSE.
'''
# get the initial gradient term so we can perform the first
# round of GD. Needed for momentum terms
grad_user, grad_item = gradient_reg(df=df,
utility=utility,
user_features=user_features,
item_features=item_features,
lmbda_a=lmbda_a,
lmbda_b=lmbda_b)
for i in range(epochs):
# update the gradient based on new feature matrices
grad_user, grad_item = gradient_reg(df=df,
utility=utility,
user_features=user_features,
item_features=item_features,
lmbda_a=lmbda_a,
lmbda_b=lmbda_b)
user_features = user_features - learning_rate*grad_user
item_features = item_features - learning_rate*grad_item
# just print out values every so often to see what is happening
# with the algo.
if(not (i+1) % 50) and (updates):
print("\niteration", i+1, ":")
print("train mse:", meanSquareError(
df, user_features, item_features))
if val is not None:
print("validation mse:", meanSquareError(
val, user_features, item_features))
if dfError is not None:
dfError = dfError.append(
[[i, meanSquareError(df, user_features, item_features)]])
# compute the final MSE
mse_train = meanSquareError(df, user_features, item_features)
# here we just check if the validation set is passed in so we can return the final cost of that as well if needed.
if val is not None:
mse_val = meanSquareError(val, user_features, item_features)
if dfError is not None:
return (user_features, item_features, mse_train, mse_val, dfError)
return (user_features, item_features, mse_train, mse_val)
if dfError is not None:
return (user_features, item_features, mse_train, dfError)
return (user_features, item_features, mse_train)
def sample_hyperparameters():
'''
This function returns a random value for each hyperparameter for MSE gradient descent.
'''
return {
"K": np.random.randint(10, 20),
"lr": np.random.normal(0.05, 0.025),
"beta": np.random.normal(0.9, 0.05),
"gamma": np.random.randint(5, 15),
"epochs": np.random.randint(50, 80)
}