disco 0.2.0 → 0.2.6
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +4 -4
- data/CHANGELOG.md +35 -0
- data/LICENSE.txt +1 -1
- data/README.md +56 -14
- data/lib/disco.rb +1 -0
- data/lib/disco/metrics.rb +10 -0
- data/lib/disco/recommender.rb +261 -93
- data/lib/disco/version.rb +1 -1
- metadata +10 -121
checksums.yaml
CHANGED
@@ -1,7 +1,7 @@
|
|
1
1
|
---
|
2
2
|
SHA256:
|
3
|
-
metadata.gz:
|
4
|
-
data.tar.gz:
|
3
|
+
metadata.gz: a7823dbe0e68967c39a59f8cdc2fe577f4366b492e0559487606b74a7de1cc84
|
4
|
+
data.tar.gz: ba40e46b203e424eccb811c6b042c9a283356c42585b7e00123b4bb2f232b1e2
|
5
5
|
SHA512:
|
6
|
-
metadata.gz:
|
7
|
-
data.tar.gz:
|
6
|
+
metadata.gz: ee43326933ac019b0bae631631ba79a7b1e03d1e9669361ef7722aa5a43b7bf2a2f49ccf8b098ab23539392fd09b83224c3cb9d340b80483179fabb45d62ee30
|
7
|
+
data.tar.gz: 9733820cc4e81b22cca51dbf89a02aa87e96cbbc1add753b2799878b5b50b549f2a27886dcfae387ad4cc158ce4bd651354f8bbd2514460ac07a60560ad5c455
|
data/CHANGELOG.md
CHANGED
@@ -1,3 +1,38 @@
|
|
1
|
+
## 0.2.6 (2021-02-24)
|
2
|
+
|
3
|
+
- Improved performance
|
4
|
+
- Improved `inspect` method
|
5
|
+
- Fixed issue with `similar_users` and `item_recs` returning the original user/item
|
6
|
+
- Fixed error with `fit` after loading
|
7
|
+
|
8
|
+
## 0.2.5 (2021-02-20)
|
9
|
+
|
10
|
+
- Added `top_items` method
|
11
|
+
- Added `optimize_similar_users` method
|
12
|
+
- Added support for Faiss for `optimize_item_recs` and `optimize_similar_users` methods
|
13
|
+
- Added `rmse` method
|
14
|
+
- Improved performance
|
15
|
+
|
16
|
+
## 0.2.4 (2021-02-15)
|
17
|
+
|
18
|
+
- Added `user_ids` and `item_ids` methods
|
19
|
+
- Added `user_id` argument to `user_factors`
|
20
|
+
- Added `item_id` argument to `item_factors`
|
21
|
+
|
22
|
+
## 0.2.3 (2020-11-28)
|
23
|
+
|
24
|
+
- Added `predict` method
|
25
|
+
- Fixed bad recommendations and scores with `user_recs` and explicit feedback
|
26
|
+
- Fixed `item_ids` option for `user_recs`
|
27
|
+
|
28
|
+
## 0.2.2 (n/a)
|
29
|
+
|
30
|
+
- Not available (released by previous gem owner)
|
31
|
+
|
32
|
+
## 0.2.1 (2020-10-28)
|
33
|
+
|
34
|
+
- Fixed issue with `user_recs` returning rated items
|
35
|
+
|
1
36
|
## 0.2.0 (2020-07-31)
|
2
37
|
|
3
38
|
- Changed score to always be between -1 and 1 for `item_recs` and `similar_users` (cosine similarity - this makes it easier to understand and consistent with `optimize_item_recs` and `optimize_similar_users`)
|
data/LICENSE.txt
CHANGED
data/README.md
CHANGED
@@ -1,12 +1,12 @@
|
|
1
1
|
# Disco
|
2
2
|
|
3
|
-
:fire:
|
3
|
+
:fire: Recommendations for Ruby and Rails using collaborative filtering
|
4
4
|
|
5
5
|
- Supports user-based and item-based recommendations
|
6
6
|
- Works with explicit and implicit feedback
|
7
7
|
- Uses high-performance matrix factorization
|
8
8
|
|
9
|
-
[![Build Status](https://
|
9
|
+
[![Build Status](https://github.com/ankane/disco/workflows/build/badge.svg?branch=master)](https://github.com/ankane/disco/actions)
|
10
10
|
|
11
11
|
## Installation
|
12
12
|
|
@@ -44,15 +44,15 @@ recommender.fit([
|
|
44
44
|
])
|
45
45
|
```
|
46
46
|
|
47
|
-
> Use `value` instead of rating for implicit feedback
|
47
|
+
> Use `value` instead of `rating` for implicit feedback
|
48
48
|
|
49
|
-
Get user-based
|
49
|
+
Get user-based recommendations - “users like you also liked”
|
50
50
|
|
51
51
|
```ruby
|
52
52
|
recommender.user_recs(user_id)
|
53
53
|
```
|
54
54
|
|
55
|
-
Get item-based
|
55
|
+
Get item-based recommendations - “users who liked this item also liked”
|
56
56
|
|
57
57
|
```ruby
|
58
58
|
recommender.item_recs(item_id)
|
@@ -64,10 +64,10 @@ Use the `count` option to specify the number of recommendations (default is 5)
|
|
64
64
|
recommender.user_recs(user_id, count: 3)
|
65
65
|
```
|
66
66
|
|
67
|
-
Get predicted ratings for specific items
|
67
|
+
Get predicted ratings for specific users and items
|
68
68
|
|
69
69
|
```ruby
|
70
|
-
recommender.
|
70
|
+
recommender.predict([{user_id: 1, item_id: 2}, {user_id: 2, item_id: 4}])
|
71
71
|
```
|
72
72
|
|
73
73
|
Get similar users
|
@@ -101,7 +101,8 @@ recommender.item_recs("Star Wars (1977)")
|
|
101
101
|
```ruby
|
102
102
|
views = Ahoy::Event.
|
103
103
|
where(name: "Viewed post").
|
104
|
-
group(:user_id
|
104
|
+
group(:user_id).
|
105
|
+
group("properties->>'post_id'"). # postgres syntax
|
105
106
|
count
|
106
107
|
|
107
108
|
data =
|
@@ -200,6 +201,8 @@ bin = File.binread("recommender.bin")
|
|
200
201
|
recommender = Marshal.load(bin)
|
201
202
|
```
|
202
203
|
|
204
|
+
Alternatively, you can store only the factors and use a library like [Neighbor](https://github.com/ankane/neighbor)
|
205
|
+
|
203
206
|
## Algorithms
|
204
207
|
|
205
208
|
Disco uses high-performance matrix factorization.
|
@@ -236,6 +239,16 @@ There are a number of ways to deal with this, but here are some common ones:
|
|
236
239
|
- For user-based recommendations, show new users the most popular items.
|
237
240
|
- For item-based recommendations, make content-based recommendations with a gem like [tf-idf-similarity](https://github.com/jpmckinney/tf-idf-similarity).
|
238
241
|
|
242
|
+
Get top items with:
|
243
|
+
|
244
|
+
```ruby
|
245
|
+
recommender = Disco::Recommender.new(top_items: true)
|
246
|
+
recommender.fit(data)
|
247
|
+
recommender.top_items
|
248
|
+
```
|
249
|
+
|
250
|
+
This uses [Wilson score](https://www.evanmiller.org/how-not-to-sort-by-average-rating.html) for explicit feedback (add [wilson_score](https://github.com/instacart/wilson_score) to your application’s Gemfile) and item frequency for implicit feedback.
|
251
|
+
|
239
252
|
## Data
|
240
253
|
|
241
254
|
Data can be an array of hashes
|
@@ -256,23 +269,29 @@ Or a Daru data frame
|
|
256
269
|
Daru::DataFrame.from_csv("ratings.csv")
|
257
270
|
```
|
258
271
|
|
259
|
-
##
|
272
|
+
## Performance
|
260
273
|
|
261
|
-
If you have a large number of users
|
274
|
+
If you have a large number of users or items, you can use an approximate nearest neighbors library like [Faiss](https://github.com/ankane/faiss) to improve the performance of certain methods.
|
262
275
|
|
263
276
|
Add this line to your application’s Gemfile:
|
264
277
|
|
265
278
|
```ruby
|
266
|
-
gem '
|
279
|
+
gem 'faiss'
|
280
|
+
```
|
281
|
+
|
282
|
+
Speed up the `user_recs` method with:
|
283
|
+
|
284
|
+
```ruby
|
285
|
+
model.optimize_user_recs
|
267
286
|
```
|
268
287
|
|
269
|
-
Speed up
|
288
|
+
Speed up the `item_recs` method with:
|
270
289
|
|
271
290
|
```ruby
|
272
291
|
model.optimize_item_recs
|
273
292
|
```
|
274
293
|
|
275
|
-
Speed up
|
294
|
+
Speed up the `similar_users` method with:
|
276
295
|
|
277
296
|
```ruby
|
278
297
|
model.optimize_similar_users
|
@@ -282,19 +301,33 @@ This should be called after fitting or loading the model.
|
|
282
301
|
|
283
302
|
## Reference
|
284
303
|
|
304
|
+
Get ids
|
305
|
+
|
306
|
+
```ruby
|
307
|
+
recommender.user_ids
|
308
|
+
recommender.item_ids
|
309
|
+
```
|
310
|
+
|
285
311
|
Get the global mean
|
286
312
|
|
287
313
|
```ruby
|
288
314
|
recommender.global_mean
|
289
315
|
```
|
290
316
|
|
291
|
-
Get
|
317
|
+
Get factors
|
292
318
|
|
293
319
|
```ruby
|
294
320
|
recommender.user_factors
|
295
321
|
recommender.item_factors
|
296
322
|
```
|
297
323
|
|
324
|
+
Get factors for specific users and items
|
325
|
+
|
326
|
+
```ruby
|
327
|
+
recommender.user_factors(user_id)
|
328
|
+
recommender.item_factors(item_id)
|
329
|
+
```
|
330
|
+
|
298
331
|
## Credits
|
299
332
|
|
300
333
|
Thanks to:
|
@@ -315,3 +348,12 @@ Everyone is encouraged to help improve this project. Here are a few ways you can
|
|
315
348
|
- Fix bugs and [submit pull requests](https://github.com/ankane/disco/pulls)
|
316
349
|
- Write, clarify, or fix documentation
|
317
350
|
- Suggest or add new features
|
351
|
+
|
352
|
+
To get started with development:
|
353
|
+
|
354
|
+
```sh
|
355
|
+
git clone https://github.com/ankane/disco.git
|
356
|
+
cd disco
|
357
|
+
bundle install
|
358
|
+
bundle exec rake test
|
359
|
+
```
|
data/lib/disco.rb
CHANGED
data/lib/disco/recommender.rb
CHANGED
@@ -1,39 +1,44 @@
|
|
1
1
|
module Disco
|
2
2
|
class Recommender
|
3
|
-
attr_reader :global_mean
|
3
|
+
attr_reader :global_mean
|
4
4
|
|
5
|
-
def initialize(factors: 8, epochs: 20, verbose: nil)
|
5
|
+
def initialize(factors: 8, epochs: 20, verbose: nil, top_items: false)
|
6
6
|
@factors = factors
|
7
7
|
@epochs = epochs
|
8
8
|
@verbose = verbose
|
9
|
+
@user_map = {}
|
10
|
+
@item_map = {}
|
11
|
+
@top_items = top_items
|
9
12
|
end
|
10
13
|
|
11
14
|
def fit(train_set, validation_set: nil)
|
12
15
|
train_set = to_dataset(train_set)
|
13
16
|
validation_set = to_dataset(validation_set) if validation_set
|
14
17
|
|
18
|
+
check_training_set(train_set)
|
19
|
+
|
20
|
+
# TODO option to set in initializer to avoid pass
|
21
|
+
# could also just check first few values
|
22
|
+
# but may be confusing if they are all missing and later ones aren't
|
15
23
|
@implicit = !train_set.any? { |v| v[:rating] }
|
16
24
|
|
25
|
+
# TODO improve performance
|
26
|
+
# (catch exception instead of checking ahead of time)
|
17
27
|
unless @implicit
|
18
|
-
|
19
|
-
check_ratings(ratings)
|
20
|
-
@min_rating = ratings.min
|
21
|
-
@max_rating = ratings.max
|
28
|
+
check_ratings(train_set)
|
22
29
|
|
23
30
|
if validation_set
|
24
|
-
check_ratings(validation_set
|
31
|
+
check_ratings(validation_set)
|
25
32
|
end
|
26
33
|
end
|
27
34
|
|
28
|
-
check_training_set(train_set)
|
29
|
-
create_maps(train_set)
|
30
|
-
|
31
35
|
@rated = Hash.new { |hash, key| hash[key] = {} }
|
32
36
|
input = []
|
33
37
|
value_key = @implicit ? :value : :rating
|
34
38
|
train_set.each do |v|
|
35
|
-
|
36
|
-
|
39
|
+
# update maps and build matrix in single pass
|
40
|
+
u = (@user_map[v[:user_id]] ||= @user_map.size)
|
41
|
+
i = (@item_map[v[:item_id]] ||= @item_map.size)
|
37
42
|
@rated[u][i] = true
|
38
43
|
|
39
44
|
# explicit will always have a value due to check_ratings
|
@@ -41,6 +46,25 @@ module Disco
|
|
41
46
|
end
|
42
47
|
@rated.default = nil
|
43
48
|
|
49
|
+
# much more efficient than checking every value in another pass
|
50
|
+
raise ArgumentError, "Missing user_id" if @user_map.key?(nil)
|
51
|
+
raise ArgumentError, "Missing item_id" if @item_map.key?(nil)
|
52
|
+
|
53
|
+
# TODO improve performance
|
54
|
+
unless @implicit
|
55
|
+
@min_rating, @max_rating = train_set.minmax_by { |o| o[:rating] }.map { |o| o[:rating] }
|
56
|
+
end
|
57
|
+
|
58
|
+
if @top_items
|
59
|
+
@item_count = [0] * @item_map.size
|
60
|
+
@item_sum = [0.0] * @item_map.size
|
61
|
+
train_set.each do |v|
|
62
|
+
i = @item_map[v[:item_id]]
|
63
|
+
@item_count[i] += 1
|
64
|
+
@item_sum[i] += (v[value_key] || 1)
|
65
|
+
end
|
66
|
+
end
|
67
|
+
|
44
68
|
eval_set = nil
|
45
69
|
if validation_set
|
46
70
|
eval_set = []
|
@@ -67,135 +91,258 @@ module Disco
|
|
67
91
|
@user_factors = model.p_factors(format: :numo)
|
68
92
|
@item_factors = model.q_factors(format: :numo)
|
69
93
|
|
70
|
-
@
|
71
|
-
@
|
94
|
+
@normalized_user_factors = nil
|
95
|
+
@normalized_item_factors = nil
|
96
|
+
|
97
|
+
@user_recs_index = nil
|
98
|
+
@similar_users_index = nil
|
99
|
+
@similar_items_index = nil
|
100
|
+
end
|
101
|
+
|
102
|
+
# generates a prediction even if a user has already rated the item
|
103
|
+
def predict(data)
|
104
|
+
data = to_dataset(data)
|
105
|
+
|
106
|
+
u = data.map { |v| @user_map[v[:user_id]] }
|
107
|
+
i = data.map { |v| @item_map[v[:item_id]] }
|
108
|
+
|
109
|
+
new_index = data.each_index.select { |index| u[index].nil? || i[index].nil? }
|
110
|
+
new_index.each do |j|
|
111
|
+
u[j] = 0
|
112
|
+
i[j] = 0
|
113
|
+
end
|
114
|
+
|
115
|
+
predictions = @user_factors[u, true].inner(@item_factors[i, true])
|
116
|
+
predictions.inplace.clip(@min_rating, @max_rating) if @min_rating
|
117
|
+
predictions[new_index] = @global_mean
|
118
|
+
predictions.to_a
|
72
119
|
end
|
73
120
|
|
74
121
|
def user_recs(user_id, count: 5, item_ids: nil)
|
122
|
+
check_fit
|
75
123
|
u = @user_map[user_id]
|
76
124
|
|
77
125
|
if u
|
78
|
-
|
79
|
-
predictions.inplace.clip(@min_rating, @max_rating) if @min_rating
|
80
|
-
|
81
|
-
predictions =
|
82
|
-
@item_map.keys.zip(predictions).map do |item_id, pred|
|
83
|
-
{item_id: item_id, score: pred}
|
84
|
-
end
|
126
|
+
rated = item_ids ? {} : @rated[u]
|
85
127
|
|
86
128
|
if item_ids
|
87
|
-
|
88
|
-
|
129
|
+
ids = Numo::NArray.cast(item_ids.map { |i| @item_map[i] }.compact)
|
130
|
+
return [] if ids.size == 0
|
131
|
+
|
132
|
+
predictions = @item_factors[ids, true].inner(@user_factors[u, true])
|
133
|
+
indexes = predictions.sort_index.reverse
|
134
|
+
indexes = indexes[0...[count + rated.size, indexes.size].min] if count
|
135
|
+
predictions = predictions[indexes]
|
136
|
+
ids = ids[indexes]
|
137
|
+
elsif @user_recs_index && count
|
138
|
+
predictions, ids = @user_recs_index.search(@user_factors[u, true].expand_dims(0), count + rated.size).map { |v| v[0, true] }
|
89
139
|
else
|
90
|
-
@
|
91
|
-
|
92
|
-
|
140
|
+
predictions = @item_factors.inner(@user_factors[u, true])
|
141
|
+
# TODO make sure reverse isn't hurting performance
|
142
|
+
indexes = predictions.sort_index.reverse
|
143
|
+
indexes = indexes[0...[count + rated.size, indexes.size].min] if count
|
144
|
+
predictions = predictions[indexes]
|
145
|
+
ids = indexes
|
93
146
|
end
|
94
147
|
|
95
|
-
predictions.
|
96
|
-
|
97
|
-
|
148
|
+
predictions.inplace.clip(@min_rating, @max_rating) if @min_rating
|
149
|
+
|
150
|
+
keys = @item_map.keys
|
151
|
+
result = []
|
152
|
+
ids.each_with_index do |item_id, i|
|
153
|
+
next if rated[item_id]
|
154
|
+
|
155
|
+
result << {item_id: keys[item_id], score: predictions[i]}
|
156
|
+
break if result.size == count
|
157
|
+
end
|
158
|
+
result
|
159
|
+
elsif @top_items
|
160
|
+
top_items(count: count)
|
98
161
|
else
|
99
|
-
# no items if user is unknown
|
100
|
-
# TODO maybe most popular items
|
101
162
|
[]
|
102
163
|
end
|
103
164
|
end
|
104
165
|
|
105
|
-
def
|
106
|
-
|
166
|
+
def similar_items(item_id, count: 5)
|
167
|
+
check_fit
|
168
|
+
similar(item_id, @item_map, normalized_item_factors, count, @similar_items_index)
|
107
169
|
end
|
108
|
-
alias_method :
|
170
|
+
alias_method :item_recs, :similar_items
|
109
171
|
|
110
|
-
def
|
111
|
-
|
172
|
+
def similar_users(user_id, count: 5)
|
173
|
+
check_fit
|
174
|
+
similar(user_id, @user_map, normalized_user_factors, count, @similar_users_index)
|
112
175
|
end
|
113
176
|
|
114
|
-
def
|
115
|
-
|
177
|
+
def top_items(count: 5)
|
178
|
+
check_fit
|
179
|
+
raise "top_items not computed" unless @top_items
|
180
|
+
|
181
|
+
if @implicit
|
182
|
+
scores = @item_count
|
183
|
+
else
|
184
|
+
require "wilson_score"
|
185
|
+
|
186
|
+
range = @min_rating..@max_rating
|
187
|
+
scores = @item_sum.zip(@item_count).map { |s, c| WilsonScore.rating_lower_bound(s / c, c, range) }
|
188
|
+
end
|
189
|
+
|
190
|
+
scores = scores.map.with_index.sort_by { |s, _| -s }
|
191
|
+
scores = scores.first(count) if count
|
192
|
+
item_ids = item_ids()
|
193
|
+
scores.map do |s, i|
|
194
|
+
{item_id: item_ids[i], score: s}
|
195
|
+
end
|
116
196
|
end
|
117
|
-
alias_method :item_recs, :similar_items
|
118
197
|
|
119
|
-
def
|
120
|
-
|
198
|
+
def user_ids
|
199
|
+
@user_map.keys
|
200
|
+
end
|
201
|
+
|
202
|
+
def item_ids
|
203
|
+
@item_map.keys
|
204
|
+
end
|
205
|
+
|
206
|
+
def user_factors(user_id = nil)
|
207
|
+
if user_id
|
208
|
+
u = @user_map[user_id]
|
209
|
+
@user_factors[u, true] if u
|
210
|
+
else
|
211
|
+
@user_factors
|
212
|
+
end
|
213
|
+
end
|
214
|
+
|
215
|
+
def item_factors(item_id = nil)
|
216
|
+
if item_id
|
217
|
+
i = @item_map[item_id]
|
218
|
+
@item_factors[i, true] if i
|
219
|
+
else
|
220
|
+
@item_factors
|
221
|
+
end
|
222
|
+
end
|
223
|
+
|
224
|
+
def optimize_user_recs
|
225
|
+
check_fit
|
226
|
+
@user_recs_index = create_index(item_factors, library: "faiss")
|
227
|
+
end
|
228
|
+
|
229
|
+
def optimize_similar_items(library: nil)
|
230
|
+
check_fit
|
231
|
+
@similar_items_index = create_index(normalized_item_factors, library: library)
|
232
|
+
end
|
233
|
+
alias_method :optimize_item_recs, :optimize_similar_items
|
234
|
+
|
235
|
+
def optimize_similar_users(library: nil)
|
236
|
+
check_fit
|
237
|
+
@similar_users_index = create_index(normalized_user_factors, library: library)
|
238
|
+
end
|
239
|
+
|
240
|
+
def inspect
|
241
|
+
to_s # for now
|
121
242
|
end
|
122
243
|
|
123
244
|
private
|
124
245
|
|
125
|
-
|
126
|
-
|
246
|
+
# factors should already be normalized for similar users/items
|
247
|
+
def create_index(factors, library:)
|
248
|
+
# TODO make Faiss the default in 0.3.0
|
249
|
+
library ||= defined?(Faiss) && !defined?(Ngt) ? "faiss" : "ngt"
|
250
|
+
|
251
|
+
case library
|
252
|
+
when "faiss"
|
253
|
+
require "faiss"
|
254
|
+
|
255
|
+
# inner product is cosine similarity with normalized vectors
|
256
|
+
# https://github.com/facebookresearch/faiss/issues/95
|
257
|
+
#
|
258
|
+
# TODO use non-exact index
|
259
|
+
# https://github.com/facebookresearch/faiss/wiki/Faiss-indexes
|
260
|
+
index = Faiss::IndexFlatIP.new(factors.shape[1])
|
261
|
+
|
262
|
+
# ids are from 0...total
|
263
|
+
# https://github.com/facebookresearch/faiss/blob/96b740abedffc8f67389f29c2a180913941534c6/faiss/Index.h#L89
|
264
|
+
index.add(factors)
|
265
|
+
|
266
|
+
index
|
267
|
+
when "ngt"
|
268
|
+
require "ngt"
|
127
269
|
|
128
|
-
|
129
|
-
|
130
|
-
|
270
|
+
# could speed up search with normalized cosine
|
271
|
+
# https://github.com/yahoojapan/NGT/issues/36
|
272
|
+
index = Ngt::Index.new(factors.shape[1], distance_type: "Cosine")
|
273
|
+
|
274
|
+
# NGT normalizes so could call create_index without normalized factors
|
275
|
+
# but keep code simple for now
|
276
|
+
ids = index.batch_insert(factors)
|
277
|
+
raise "Unexpected ids. Please report a bug." if ids.first != 1 || ids.last != factors.shape[0]
|
278
|
+
|
279
|
+
index
|
280
|
+
else
|
281
|
+
raise ArgumentError, "Invalid library: #{library}"
|
282
|
+
end
|
131
283
|
end
|
132
284
|
|
133
|
-
def
|
134
|
-
@
|
285
|
+
def normalized_user_factors
|
286
|
+
@normalized_user_factors ||= normalize(@user_factors)
|
135
287
|
end
|
136
288
|
|
137
|
-
def
|
138
|
-
@
|
289
|
+
def normalized_item_factors
|
290
|
+
@normalized_item_factors ||= normalize(@item_factors)
|
139
291
|
end
|
140
292
|
|
141
|
-
def
|
293
|
+
def normalize(factors)
|
142
294
|
norms = Numo::SFloat::Math.sqrt((factors * factors).sum(axis: 1))
|
143
295
|
norms[norms.eq(0)] = 1e-10 # no zeros
|
144
|
-
norms
|
296
|
+
factors / norms.expand_dims(1)
|
145
297
|
end
|
146
298
|
|
147
|
-
def similar(id, map,
|
299
|
+
def similar(id, map, norm_factors, count, index)
|
148
300
|
i = map[id]
|
149
|
-
|
301
|
+
|
302
|
+
if i && norm_factors.shape[0] > 1
|
150
303
|
if index && count
|
151
|
-
|
152
|
-
|
153
|
-
|
154
|
-
|
155
|
-
|
156
|
-
|
157
|
-
|
158
|
-
|
159
|
-
}
|
304
|
+
if defined?(Faiss) && index.is_a?(Faiss::Index)
|
305
|
+
predictions, ids = index.search(norm_factors[i, true].expand_dims(0), count + 1).map { |v| v.to_a[0] }
|
306
|
+
else
|
307
|
+
result = index.search(norm_factors[i, true], size: count + 1)
|
308
|
+
# ids from batch_insert start at 1 instead of 0
|
309
|
+
ids = result.map { |v| v[:id] - 1 }
|
310
|
+
# convert cosine distance to cosine similarity
|
311
|
+
predictions = result.map { |v| 1 - v[:distance] }
|
160
312
|
end
|
161
313
|
else
|
162
|
-
predictions =
|
163
|
-
|
164
|
-
|
165
|
-
|
166
|
-
|
167
|
-
end
|
168
|
-
|
169
|
-
max_score = predictions.delete_at(i)[:score]
|
170
|
-
predictions.sort_by! { |pred| -pred[:score] } # already sorted by id
|
171
|
-
predictions = predictions.first(count) if count
|
172
|
-
# divide by max score to get cosine similarity
|
173
|
-
# only need to do for returned records
|
174
|
-
predictions.each { |pred| pred[:score] /= max_score }
|
175
|
-
predictions
|
314
|
+
predictions = norm_factors.inner(norm_factors[i, true])
|
315
|
+
indexes = predictions.sort_index.reverse
|
316
|
+
indexes = indexes[0...[count + 1, indexes.size].min] if count
|
317
|
+
predictions = predictions[indexes]
|
318
|
+
ids = indexes
|
176
319
|
end
|
177
|
-
else
|
178
|
-
[]
|
179
|
-
end
|
180
|
-
end
|
181
320
|
|
182
|
-
|
183
|
-
|
184
|
-
|
321
|
+
keys = map.keys
|
322
|
+
|
323
|
+
# TODO use user_id for similar_users in 0.3.0
|
324
|
+
key = :item_id
|
185
325
|
|
186
|
-
|
187
|
-
|
326
|
+
result = []
|
327
|
+
# items can have the same score
|
328
|
+
# so original item may not be at index 0
|
329
|
+
ids.each_with_index do |id, j|
|
330
|
+
next if id == i
|
188
331
|
|
189
|
-
|
190
|
-
|
332
|
+
result << {key => keys[id], score: predictions[j]}
|
333
|
+
end
|
334
|
+
result
|
335
|
+
else
|
336
|
+
[]
|
337
|
+
end
|
191
338
|
end
|
192
339
|
|
193
340
|
def check_ratings(ratings)
|
194
|
-
unless ratings.all? { |r| !r.nil? }
|
195
|
-
raise ArgumentError, "Missing
|
341
|
+
unless ratings.all? { |r| !r[:rating].nil? }
|
342
|
+
raise ArgumentError, "Missing rating"
|
196
343
|
end
|
197
|
-
unless ratings.all? { |r| r.is_a?(Numeric) }
|
198
|
-
raise ArgumentError, "
|
344
|
+
unless ratings.all? { |r| r[:rating].is_a?(Numeric) }
|
345
|
+
raise ArgumentError, "Rating must be numeric"
|
199
346
|
end
|
200
347
|
end
|
201
348
|
|
@@ -203,6 +350,10 @@ module Disco
|
|
203
350
|
raise ArgumentError, "No training data" if train_set.empty?
|
204
351
|
end
|
205
352
|
|
353
|
+
def check_fit
|
354
|
+
raise "Not fit" unless defined?(@implicit)
|
355
|
+
end
|
356
|
+
|
206
357
|
def to_dataset(dataset)
|
207
358
|
if defined?(Rover::DataFrame) && dataset.is_a?(Rover::DataFrame)
|
208
359
|
# convert keys to symbols
|
@@ -230,7 +381,10 @@ module Disco
|
|
230
381
|
rated: @rated,
|
231
382
|
global_mean: @global_mean,
|
232
383
|
user_factors: @user_factors,
|
233
|
-
item_factors: @item_factors
|
384
|
+
item_factors: @item_factors,
|
385
|
+
factors: @factors,
|
386
|
+
epochs: @epochs,
|
387
|
+
verbose: @verbose
|
234
388
|
}
|
235
389
|
|
236
390
|
unless @implicit
|
@@ -238,6 +392,11 @@ module Disco
|
|
238
392
|
obj[:max_rating] = @max_rating
|
239
393
|
end
|
240
394
|
|
395
|
+
if @top_items
|
396
|
+
obj[:item_count] = @item_count
|
397
|
+
obj[:item_sum] = @item_sum
|
398
|
+
end
|
399
|
+
|
241
400
|
obj
|
242
401
|
end
|
243
402
|
|
@@ -249,11 +408,20 @@ module Disco
|
|
249
408
|
@global_mean = obj[:global_mean]
|
250
409
|
@user_factors = obj[:user_factors]
|
251
410
|
@item_factors = obj[:item_factors]
|
411
|
+
@factors = obj[:factors]
|
412
|
+
@epochs = obj[:epochs]
|
413
|
+
@verbose = obj[:verbose]
|
252
414
|
|
253
415
|
unless @implicit
|
254
416
|
@min_rating = obj[:min_rating]
|
255
417
|
@max_rating = obj[:max_rating]
|
256
418
|
end
|
419
|
+
|
420
|
+
@top_items = obj.key?(:item_count)
|
421
|
+
if @top_items
|
422
|
+
@item_count = obj[:item_count]
|
423
|
+
@item_sum = obj[:item_sum]
|
424
|
+
end
|
257
425
|
end
|
258
426
|
end
|
259
427
|
end
|
data/lib/disco/version.rb
CHANGED
metadata
CHANGED
@@ -1,14 +1,14 @@
|
|
1
1
|
--- !ruby/object:Gem::Specification
|
2
2
|
name: disco
|
3
3
|
version: !ruby/object:Gem::Version
|
4
|
-
version: 0.2.
|
4
|
+
version: 0.2.6
|
5
5
|
platform: ruby
|
6
6
|
authors:
|
7
7
|
- Andrew Kane
|
8
|
-
autorequire:
|
8
|
+
autorequire:
|
9
9
|
bindir: bin
|
10
10
|
cert_chain: []
|
11
|
-
date:
|
11
|
+
date: 2021-02-24 00:00:00.000000000 Z
|
12
12
|
dependencies:
|
13
13
|
- !ruby/object:Gem::Dependency
|
14
14
|
name: libmf
|
@@ -38,120 +38,8 @@ dependencies:
|
|
38
38
|
- - ">="
|
39
39
|
- !ruby/object:Gem::Version
|
40
40
|
version: '0'
|
41
|
-
|
42
|
-
|
43
|
-
requirement: !ruby/object:Gem::Requirement
|
44
|
-
requirements:
|
45
|
-
- - ">="
|
46
|
-
- !ruby/object:Gem::Version
|
47
|
-
version: '0'
|
48
|
-
type: :development
|
49
|
-
prerelease: false
|
50
|
-
version_requirements: !ruby/object:Gem::Requirement
|
51
|
-
requirements:
|
52
|
-
- - ">="
|
53
|
-
- !ruby/object:Gem::Version
|
54
|
-
version: '0'
|
55
|
-
- !ruby/object:Gem::Dependency
|
56
|
-
name: rake
|
57
|
-
requirement: !ruby/object:Gem::Requirement
|
58
|
-
requirements:
|
59
|
-
- - ">="
|
60
|
-
- !ruby/object:Gem::Version
|
61
|
-
version: '0'
|
62
|
-
type: :development
|
63
|
-
prerelease: false
|
64
|
-
version_requirements: !ruby/object:Gem::Requirement
|
65
|
-
requirements:
|
66
|
-
- - ">="
|
67
|
-
- !ruby/object:Gem::Version
|
68
|
-
version: '0'
|
69
|
-
- !ruby/object:Gem::Dependency
|
70
|
-
name: minitest
|
71
|
-
requirement: !ruby/object:Gem::Requirement
|
72
|
-
requirements:
|
73
|
-
- - ">="
|
74
|
-
- !ruby/object:Gem::Version
|
75
|
-
version: '5'
|
76
|
-
type: :development
|
77
|
-
prerelease: false
|
78
|
-
version_requirements: !ruby/object:Gem::Requirement
|
79
|
-
requirements:
|
80
|
-
- - ">="
|
81
|
-
- !ruby/object:Gem::Version
|
82
|
-
version: '5'
|
83
|
-
- !ruby/object:Gem::Dependency
|
84
|
-
name: activerecord
|
85
|
-
requirement: !ruby/object:Gem::Requirement
|
86
|
-
requirements:
|
87
|
-
- - ">="
|
88
|
-
- !ruby/object:Gem::Version
|
89
|
-
version: '0'
|
90
|
-
type: :development
|
91
|
-
prerelease: false
|
92
|
-
version_requirements: !ruby/object:Gem::Requirement
|
93
|
-
requirements:
|
94
|
-
- - ">="
|
95
|
-
- !ruby/object:Gem::Version
|
96
|
-
version: '0'
|
97
|
-
- !ruby/object:Gem::Dependency
|
98
|
-
name: sqlite3
|
99
|
-
requirement: !ruby/object:Gem::Requirement
|
100
|
-
requirements:
|
101
|
-
- - ">="
|
102
|
-
- !ruby/object:Gem::Version
|
103
|
-
version: '0'
|
104
|
-
type: :development
|
105
|
-
prerelease: false
|
106
|
-
version_requirements: !ruby/object:Gem::Requirement
|
107
|
-
requirements:
|
108
|
-
- - ">="
|
109
|
-
- !ruby/object:Gem::Version
|
110
|
-
version: '0'
|
111
|
-
- !ruby/object:Gem::Dependency
|
112
|
-
name: daru
|
113
|
-
requirement: !ruby/object:Gem::Requirement
|
114
|
-
requirements:
|
115
|
-
- - ">="
|
116
|
-
- !ruby/object:Gem::Version
|
117
|
-
version: '0'
|
118
|
-
type: :development
|
119
|
-
prerelease: false
|
120
|
-
version_requirements: !ruby/object:Gem::Requirement
|
121
|
-
requirements:
|
122
|
-
- - ">="
|
123
|
-
- !ruby/object:Gem::Version
|
124
|
-
version: '0'
|
125
|
-
- !ruby/object:Gem::Dependency
|
126
|
-
name: rover-df
|
127
|
-
requirement: !ruby/object:Gem::Requirement
|
128
|
-
requirements:
|
129
|
-
- - ">="
|
130
|
-
- !ruby/object:Gem::Version
|
131
|
-
version: '0'
|
132
|
-
type: :development
|
133
|
-
prerelease: false
|
134
|
-
version_requirements: !ruby/object:Gem::Requirement
|
135
|
-
requirements:
|
136
|
-
- - ">="
|
137
|
-
- !ruby/object:Gem::Version
|
138
|
-
version: '0'
|
139
|
-
- !ruby/object:Gem::Dependency
|
140
|
-
name: ngt
|
141
|
-
requirement: !ruby/object:Gem::Requirement
|
142
|
-
requirements:
|
143
|
-
- - ">="
|
144
|
-
- !ruby/object:Gem::Version
|
145
|
-
version: 0.3.0
|
146
|
-
type: :development
|
147
|
-
prerelease: false
|
148
|
-
version_requirements: !ruby/object:Gem::Requirement
|
149
|
-
requirements:
|
150
|
-
- - ">="
|
151
|
-
- !ruby/object:Gem::Version
|
152
|
-
version: 0.3.0
|
153
|
-
description:
|
154
|
-
email: andrew@chartkick.com
|
41
|
+
description:
|
42
|
+
email: andrew@ankane.org
|
155
43
|
executables: []
|
156
44
|
extensions: []
|
157
45
|
extra_rdoc_files: []
|
@@ -163,6 +51,7 @@ files:
|
|
163
51
|
- lib/disco.rb
|
164
52
|
- lib/disco/data.rb
|
165
53
|
- lib/disco/engine.rb
|
54
|
+
- lib/disco/metrics.rb
|
166
55
|
- lib/disco/model.rb
|
167
56
|
- lib/disco/recommender.rb
|
168
57
|
- lib/disco/version.rb
|
@@ -172,7 +61,7 @@ homepage: https://github.com/ankane/disco
|
|
172
61
|
licenses:
|
173
62
|
- MIT
|
174
63
|
metadata: {}
|
175
|
-
post_install_message:
|
64
|
+
post_install_message:
|
176
65
|
rdoc_options: []
|
177
66
|
require_paths:
|
178
67
|
- lib
|
@@ -187,8 +76,8 @@ required_rubygems_version: !ruby/object:Gem::Requirement
|
|
187
76
|
- !ruby/object:Gem::Version
|
188
77
|
version: '0'
|
189
78
|
requirements: []
|
190
|
-
rubygems_version: 3.
|
191
|
-
signing_key:
|
79
|
+
rubygems_version: 3.2.3
|
80
|
+
signing_key:
|
192
81
|
specification_version: 4
|
193
|
-
summary:
|
82
|
+
summary: Recommendations for Ruby and Rails using collaborative filtering
|
194
83
|
test_files: []
|