disco 0.2.5 → 0.2.8
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +4 -4
- data/CHANGELOG.md +15 -0
- data/README.md +36 -19
- data/lib/disco/model.rb +1 -0
- data/lib/disco/recommender.rb +89 -45
- data/lib/disco/version.rb +1 -1
- metadata +3 -3
checksums.yaml
CHANGED
@@ -1,7 +1,7 @@
|
|
1
1
|
---
|
2
2
|
SHA256:
|
3
|
-
metadata.gz:
|
4
|
-
data.tar.gz:
|
3
|
+
metadata.gz: 0ec370f448c5cc8aebb2860b0580466687874eb34c165e1a3d0254a1c6e701d7
|
4
|
+
data.tar.gz: 1b080c37206371ee59ce184ae420c5fa1de60da0714ef17ea9459b31fcdd22ab
|
5
5
|
SHA512:
|
6
|
-
metadata.gz:
|
7
|
-
data.tar.gz:
|
6
|
+
metadata.gz: 1802e0fbf68ee489891f94e468c3b24df0eb8463de4e55a7f50a9dbc86bda7631ca15d5b84d0d584dbd341e994db84a82f00f6753c70eda4a7400376f0443df5
|
7
|
+
data.tar.gz: bee0645357a5fc4eb226e75d4b56208e7377952911dd19ccad8f3072eee3eaf4d4a318933dcb4ec4d75273630e278dbf98b3ac21f5987baea90261d15cc2d851
|
data/CHANGELOG.md
CHANGED
@@ -1,3 +1,18 @@
|
|
1
|
+
## 0.2.8 (2022-03-13)
|
2
|
+
|
3
|
+
- Fixed error with `top_items` with all same rating
|
4
|
+
|
5
|
+
## 0.2.7 (2021-08-06)
|
6
|
+
|
7
|
+
- Added warning for `value`
|
8
|
+
|
9
|
+
## 0.2.6 (2021-02-24)
|
10
|
+
|
11
|
+
- Improved performance
|
12
|
+
- Improved `inspect` method
|
13
|
+
- Fixed issue with `similar_users` and `item_recs` returning the original user/item
|
14
|
+
- Fixed error with `fit` after loading
|
15
|
+
|
1
16
|
## 0.2.5 (2021-02-20)
|
2
17
|
|
3
18
|
- Added `top_items` method
|
data/README.md
CHANGED
@@ -35,16 +35,16 @@ recommender.fit([
|
|
35
35
|
|
36
36
|
> IDs can be integers, strings, or any other data type
|
37
37
|
|
38
|
-
If users don’t rate items directly (for instance, they’re purchasing items or reading posts), this is known as implicit feedback. Leave out the rating
|
38
|
+
If users don’t rate items directly (for instance, they’re purchasing items or reading posts), this is known as implicit feedback. Leave out the rating.
|
39
39
|
|
40
40
|
```ruby
|
41
41
|
recommender.fit([
|
42
|
-
{user_id: 1, item_id: 1
|
43
|
-
{user_id: 2, item_id: 1
|
42
|
+
{user_id: 1, item_id: 1},
|
43
|
+
{user_id: 2, item_id: 1}
|
44
44
|
])
|
45
45
|
```
|
46
46
|
|
47
|
-
>
|
47
|
+
> Each `user_id`/`item_id` combination should only appear once
|
48
48
|
|
49
49
|
Get user-based recommendations - “users like you also liked”
|
50
50
|
|
@@ -99,18 +99,13 @@ recommender.item_recs("Star Wars (1977)")
|
|
99
99
|
[Ahoy](https://github.com/ankane/ahoy) is a great source for implicit feedback
|
100
100
|
|
101
101
|
```ruby
|
102
|
-
views = Ahoy::Event.
|
103
|
-
where(name: "Viewed post").
|
104
|
-
group(:user_id).
|
105
|
-
group("properties->>'post_id'"). # postgres syntax
|
106
|
-
count
|
102
|
+
views = Ahoy::Event.where(name: "Viewed post").group(:user_id).group_prop(:post_id).count
|
107
103
|
|
108
104
|
data =
|
109
|
-
views.map do |(user_id, post_id),
|
105
|
+
views.map do |(user_id, post_id), _|
|
110
106
|
{
|
111
107
|
user_id: user_id,
|
112
|
-
item_id: post_id
|
113
|
-
value: count
|
108
|
+
item_id: post_id
|
114
109
|
}
|
115
110
|
end
|
116
111
|
```
|
@@ -201,7 +196,7 @@ bin = File.binread("recommender.bin")
|
|
201
196
|
recommender = Marshal.load(bin)
|
202
197
|
```
|
203
198
|
|
204
|
-
Alternatively, you can store only the factors and use a library like [Neighbor](https://github.com/ankane/neighbor)
|
199
|
+
Alternatively, you can store only the factors and use a library like [Neighbor](https://github.com/ankane/neighbor). See the [examples](https://github.com/ankane/neighbor/tree/master/examples).
|
205
200
|
|
206
201
|
## Algorithms
|
207
202
|
|
@@ -247,7 +242,7 @@ recommender.fit(data)
|
|
247
242
|
recommender.top_items
|
248
243
|
```
|
249
244
|
|
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) your application’s Gemfile) and item frequency for implicit feedback.
|
245
|
+
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
246
|
|
252
247
|
## Data
|
253
248
|
|
@@ -269,7 +264,7 @@ Or a Daru data frame
|
|
269
264
|
Daru::DataFrame.from_csv("ratings.csv")
|
270
265
|
```
|
271
266
|
|
272
|
-
## Performance
|
267
|
+
## Performance
|
273
268
|
|
274
269
|
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.
|
275
270
|
|
@@ -282,22 +277,22 @@ gem 'faiss'
|
|
282
277
|
Speed up the `user_recs` method with:
|
283
278
|
|
284
279
|
```ruby
|
285
|
-
|
280
|
+
recommender.optimize_user_recs
|
286
281
|
```
|
287
282
|
|
288
283
|
Speed up the `item_recs` method with:
|
289
284
|
|
290
285
|
```ruby
|
291
|
-
|
286
|
+
recommender.optimize_item_recs
|
292
287
|
```
|
293
288
|
|
294
289
|
Speed up the `similar_users` method with:
|
295
290
|
|
296
291
|
```ruby
|
297
|
-
|
292
|
+
recommender.optimize_similar_users
|
298
293
|
```
|
299
294
|
|
300
|
-
This should be called after fitting or loading the
|
295
|
+
This should be called after fitting or loading the recommender.
|
301
296
|
|
302
297
|
## Reference
|
303
298
|
|
@@ -336,6 +331,28 @@ Thanks to:
|
|
336
331
|
- [Implicit](https://github.com/benfred/implicit/) for serving as an initial reference for user and item similarity
|
337
332
|
- [@dasch](https://github.com/dasch) for the gem name
|
338
333
|
|
334
|
+
## Upgrading
|
335
|
+
|
336
|
+
### 0.2.7
|
337
|
+
|
338
|
+
There’s now a warning when passing `:value` with implicit feedback, as this has no effect on recommendations and can be removed. Earlier versions of the library incorrectly stated this was used.
|
339
|
+
|
340
|
+
```ruby
|
341
|
+
recommender.fit([
|
342
|
+
{user_id: 1, item_id: 1, value: 1},
|
343
|
+
{user_id: 2, item_id: 1, value: 3}
|
344
|
+
])
|
345
|
+
```
|
346
|
+
|
347
|
+
to:
|
348
|
+
|
349
|
+
```ruby
|
350
|
+
recommender.fit([
|
351
|
+
{user_id: 1, item_id: 1},
|
352
|
+
{user_id: 2, item_id: 1}
|
353
|
+
])
|
354
|
+
```
|
355
|
+
|
339
356
|
## History
|
340
357
|
|
341
358
|
View the [changelog](https://github.com/ankane/disco/blob/master/CHANGELOG.md)
|
data/lib/disco/model.rb
CHANGED
@@ -10,6 +10,7 @@ module Disco
|
|
10
10
|
|
11
11
|
has_many :"recommended_#{name}", -> { where("disco_recommendations.context = ?", name).order("disco_recommendations.score DESC") }, through: :recommendations, source: :item, source_type: class_name
|
12
12
|
|
13
|
+
# TODO use fetch for item_id and score in 0.3.0
|
13
14
|
define_method("update_recommended_#{name}") do |items|
|
14
15
|
now = Time.now
|
15
16
|
items = items.map { |item| {subject_type: model_name.name, subject_id: id, item_type: class_name, item_id: item[:item_id], context: name, score: item[:score], created_at: now, updated_at: now} }
|
data/lib/disco/recommender.rb
CHANGED
@@ -17,38 +17,54 @@ module Disco
|
|
17
17
|
|
18
18
|
check_training_set(train_set)
|
19
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
|
20
23
|
@implicit = !train_set.any? { |v| v[:rating] }
|
24
|
+
|
25
|
+
if @implicit && train_set.any? { |v| v[:value] }
|
26
|
+
warn "[disco] WARNING: Passing `:value` with implicit feedback has no effect on recommendations and can be removed. Earlier versions of the library incorrectly stated this was used."
|
27
|
+
end
|
28
|
+
|
29
|
+
# TODO improve performance
|
30
|
+
# (catch exception instead of checking ahead of time)
|
21
31
|
unless @implicit
|
22
32
|
check_ratings(train_set)
|
23
|
-
@min_rating, @max_rating = train_set.minmax_by { |o| o[:rating] }.map { |o| o[:rating] }
|
24
33
|
|
25
34
|
if validation_set
|
26
35
|
check_ratings(validation_set)
|
27
36
|
end
|
28
37
|
end
|
29
38
|
|
30
|
-
update_maps(train_set)
|
31
|
-
|
32
39
|
@rated = Hash.new { |hash, key| hash[key] = {} }
|
33
40
|
input = []
|
34
|
-
value_key = @implicit ? :value : :rating
|
35
41
|
train_set.each do |v|
|
36
|
-
|
37
|
-
|
42
|
+
# update maps and build matrix in single pass
|
43
|
+
u = (@user_map[v[:user_id]] ||= @user_map.size)
|
44
|
+
i = (@item_map[v[:item_id]] ||= @item_map.size)
|
38
45
|
@rated[u][i] = true
|
39
46
|
|
40
47
|
# explicit will always have a value due to check_ratings
|
41
|
-
input << [u, i,
|
48
|
+
input << [u, i, @implicit ? 1 : v[:rating]]
|
42
49
|
end
|
43
50
|
@rated.default = nil
|
44
51
|
|
52
|
+
# much more efficient than checking every value in another pass
|
53
|
+
raise ArgumentError, "Missing user_id" if @user_map.key?(nil)
|
54
|
+
raise ArgumentError, "Missing item_id" if @item_map.key?(nil)
|
55
|
+
|
56
|
+
# TODO improve performance
|
57
|
+
unless @implicit
|
58
|
+
@min_rating, @max_rating = train_set.minmax_by { |o| o[:rating] }.map { |o| o[:rating] }
|
59
|
+
end
|
60
|
+
|
45
61
|
if @top_items
|
46
62
|
@item_count = [0] * @item_map.size
|
47
63
|
@item_sum = [0.0] * @item_map.size
|
48
64
|
train_set.each do |v|
|
49
65
|
i = @item_map[v[:item_id]]
|
50
66
|
@item_count[i] += 1
|
51
|
-
@item_sum[i] += (v[
|
67
|
+
@item_sum[i] += (@implicit ? 1 : v[:rating])
|
52
68
|
end
|
53
69
|
end
|
54
70
|
|
@@ -63,7 +79,7 @@ module Disco
|
|
63
79
|
u ||= -1
|
64
80
|
i ||= -1
|
65
81
|
|
66
|
-
eval_set << [u, i,
|
82
|
+
eval_set << [u, i, @implicit ? 1 : v[:rating]]
|
67
83
|
end
|
68
84
|
end
|
69
85
|
|
@@ -78,6 +94,9 @@ module Disco
|
|
78
94
|
@user_factors = model.p_factors(format: :numo)
|
79
95
|
@item_factors = model.q_factors(format: :numo)
|
80
96
|
|
97
|
+
@normalized_user_factors = nil
|
98
|
+
@normalized_item_factors = nil
|
99
|
+
|
81
100
|
@user_recs_index = nil
|
82
101
|
@similar_users_index = nil
|
83
102
|
@similar_items_index = nil
|
@@ -122,8 +141,7 @@ module Disco
|
|
122
141
|
predictions, ids = @user_recs_index.search(@user_factors[u, true].expand_dims(0), count + rated.size).map { |v| v[0, true] }
|
123
142
|
else
|
124
143
|
predictions = @item_factors.inner(@user_factors[u, true])
|
125
|
-
|
126
|
-
indexes = predictions.sort_index.reverse
|
144
|
+
indexes = predictions.sort_index.reverse # reverse just creates view
|
127
145
|
indexes = indexes[0...[count + rated.size, indexes.size].min] if count
|
128
146
|
predictions = predictions[indexes]
|
129
147
|
ids = indexes
|
@@ -149,13 +167,13 @@ module Disco
|
|
149
167
|
|
150
168
|
def similar_items(item_id, count: 5)
|
151
169
|
check_fit
|
152
|
-
similar(item_id, @item_map,
|
170
|
+
similar(item_id, @item_map, normalized_item_factors, count, @similar_items_index)
|
153
171
|
end
|
154
172
|
alias_method :item_recs, :similar_items
|
155
173
|
|
156
174
|
def similar_users(user_id, count: 5)
|
157
175
|
check_fit
|
158
|
-
similar(user_id, @user_map,
|
176
|
+
similar(user_id, @user_map, normalized_user_factors, count, @similar_users_index)
|
159
177
|
end
|
160
178
|
|
161
179
|
def top_items(count: 5)
|
@@ -163,19 +181,38 @@ module Disco
|
|
163
181
|
raise "top_items not computed" unless @top_items
|
164
182
|
|
165
183
|
if @implicit
|
166
|
-
scores = @item_count
|
184
|
+
scores = Numo::UInt64.cast(@item_count)
|
167
185
|
else
|
168
186
|
require "wilson_score"
|
169
187
|
|
170
|
-
range =
|
171
|
-
|
188
|
+
range =
|
189
|
+
if @min_rating == @max_rating
|
190
|
+
# TODO remove temp fix
|
191
|
+
(@min_rating - 1)..@max_rating
|
192
|
+
else
|
193
|
+
@min_rating..@max_rating
|
194
|
+
end
|
195
|
+
scores = Numo::DFloat.cast(@item_sum.zip(@item_count).map { |s, c| WilsonScore.rating_lower_bound(s / c, c, range) })
|
196
|
+
|
197
|
+
# TODO uncomment in 0.3.0
|
198
|
+
# wilson score with continuity correction
|
199
|
+
# https://en.wikipedia.org/wiki/Binomial_proportion_confidence_interval#Wilson_score_interval_with_continuity_correction
|
200
|
+
# z = 1.96 # 95% confidence
|
201
|
+
# range = @max_rating - @min_rating
|
202
|
+
# n = Numo::DFloat.cast(@item_count)
|
203
|
+
# phat = (Numo::DFloat.cast(@item_sum) - (@min_rating * n)) / range / n
|
204
|
+
# phat = (phat - (1 / (2 * n))).clip(0, nil) # continuity correction
|
205
|
+
# scores = (phat + z**2 / (2 * n) - z * Numo::DFloat::Math.sqrt((phat * (1 - phat) + z**2 / (4 * n)) / n)) / (1 + z**2 / n)
|
206
|
+
# scores = scores * range + @min_rating
|
172
207
|
end
|
173
208
|
|
174
|
-
|
175
|
-
|
176
|
-
|
177
|
-
|
178
|
-
|
209
|
+
indexes = scores.sort_index.reverse
|
210
|
+
indexes = indexes[0...[count, indexes.size].min] if count
|
211
|
+
scores = scores[indexes]
|
212
|
+
|
213
|
+
keys = @item_map.keys
|
214
|
+
indexes.size.times.map do |i|
|
215
|
+
{item_id: keys[indexes[i]], score: scores[i]}
|
179
216
|
end
|
180
217
|
end
|
181
218
|
|
@@ -212,13 +249,17 @@ module Disco
|
|
212
249
|
|
213
250
|
def optimize_similar_items(library: nil)
|
214
251
|
check_fit
|
215
|
-
@similar_items_index = create_index(
|
252
|
+
@similar_items_index = create_index(normalized_item_factors, library: library)
|
216
253
|
end
|
217
254
|
alias_method :optimize_item_recs, :optimize_similar_items
|
218
255
|
|
219
256
|
def optimize_similar_users(library: nil)
|
220
257
|
check_fit
|
221
|
-
@similar_users_index = create_index(
|
258
|
+
@similar_users_index = create_index(normalized_user_factors, library: library)
|
259
|
+
end
|
260
|
+
|
261
|
+
def inspect
|
262
|
+
to_s # for now
|
222
263
|
end
|
223
264
|
|
224
265
|
private
|
@@ -235,8 +276,9 @@ module Disco
|
|
235
276
|
# inner product is cosine similarity with normalized vectors
|
236
277
|
# https://github.com/facebookresearch/faiss/issues/95
|
237
278
|
#
|
238
|
-
# TODO use non-exact index
|
279
|
+
# TODO use non-exact index in 0.3.0
|
239
280
|
# https://github.com/facebookresearch/faiss/wiki/Faiss-indexes
|
281
|
+
# index = Faiss::IndexHNSWFlat.new(factors.shape[1], 32, :inner_product)
|
240
282
|
index = Faiss::IndexFlatIP.new(factors.shape[1])
|
241
283
|
|
242
284
|
# ids are from 0...total
|
@@ -251,7 +293,7 @@ module Disco
|
|
251
293
|
# https://github.com/yahoojapan/NGT/issues/36
|
252
294
|
index = Ngt::Index.new(factors.shape[1], distance_type: "Cosine")
|
253
295
|
|
254
|
-
# NGT normalizes so could call create_index
|
296
|
+
# NGT normalizes so could call create_index without normalized factors
|
255
297
|
# but keep code simple for now
|
256
298
|
ids = index.batch_insert(factors)
|
257
299
|
raise "Unexpected ids. Please report a bug." if ids.first != 1 || ids.last != factors.shape[0]
|
@@ -262,15 +304,15 @@ module Disco
|
|
262
304
|
end
|
263
305
|
end
|
264
306
|
|
265
|
-
def
|
266
|
-
@
|
307
|
+
def normalized_user_factors
|
308
|
+
@normalized_user_factors ||= normalize(@user_factors)
|
267
309
|
end
|
268
310
|
|
269
|
-
def
|
270
|
-
@
|
311
|
+
def normalized_item_factors
|
312
|
+
@normalized_item_factors ||= normalize(@item_factors)
|
271
313
|
end
|
272
314
|
|
273
|
-
def
|
315
|
+
def normalize(factors)
|
274
316
|
norms = Numo::SFloat::Math.sqrt((factors * factors).sum(axis: 1))
|
275
317
|
norms[norms.eq(0)] = 1e-10 # no zeros
|
276
318
|
factors / norms.expand_dims(1)
|
@@ -303,30 +345,26 @@ module Disco
|
|
303
345
|
# TODO use user_id for similar_users in 0.3.0
|
304
346
|
key = :item_id
|
305
347
|
|
306
|
-
|
307
|
-
|
348
|
+
result = []
|
349
|
+
# items can have the same score
|
350
|
+
# so original item may not be at index 0
|
351
|
+
ids.each_with_index do |id, j|
|
352
|
+
next if id == i
|
353
|
+
|
354
|
+
result << {key => keys[id], score: predictions[j]}
|
308
355
|
end
|
356
|
+
result
|
309
357
|
else
|
310
358
|
[]
|
311
359
|
end
|
312
360
|
end
|
313
361
|
|
314
|
-
def update_maps(train_set)
|
315
|
-
raise ArgumentError, "Missing user_id" if train_set.any? { |v| v[:user_id].nil? }
|
316
|
-
raise ArgumentError, "Missing item_id" if train_set.any? { |v| v[:item_id].nil? }
|
317
|
-
|
318
|
-
train_set.each do |v|
|
319
|
-
@user_map[v[:user_id]] ||= @user_map.size
|
320
|
-
@item_map[v[:item_id]] ||= @item_map.size
|
321
|
-
end
|
322
|
-
end
|
323
|
-
|
324
362
|
def check_ratings(ratings)
|
325
363
|
unless ratings.all? { |r| !r[:rating].nil? }
|
326
|
-
raise ArgumentError, "Missing
|
364
|
+
raise ArgumentError, "Missing rating"
|
327
365
|
end
|
328
366
|
unless ratings.all? { |r| r[:rating].is_a?(Numeric) }
|
329
|
-
raise ArgumentError, "
|
367
|
+
raise ArgumentError, "Rating must be numeric"
|
330
368
|
end
|
331
369
|
end
|
332
370
|
|
@@ -365,7 +403,10 @@ module Disco
|
|
365
403
|
rated: @rated,
|
366
404
|
global_mean: @global_mean,
|
367
405
|
user_factors: @user_factors,
|
368
|
-
item_factors: @item_factors
|
406
|
+
item_factors: @item_factors,
|
407
|
+
factors: @factors,
|
408
|
+
epochs: @epochs,
|
409
|
+
verbose: @verbose
|
369
410
|
}
|
370
411
|
|
371
412
|
unless @implicit
|
@@ -389,6 +430,9 @@ module Disco
|
|
389
430
|
@global_mean = obj[:global_mean]
|
390
431
|
@user_factors = obj[:user_factors]
|
391
432
|
@item_factors = obj[:item_factors]
|
433
|
+
@factors = obj[:factors]
|
434
|
+
@epochs = obj[:epochs]
|
435
|
+
@verbose = obj[:verbose]
|
392
436
|
|
393
437
|
unless @implicit
|
394
438
|
@min_rating = obj[:min_rating]
|
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: 0.2.8
|
5
5
|
platform: ruby
|
6
6
|
authors:
|
7
7
|
- Andrew Kane
|
8
8
|
autorequire:
|
9
9
|
bindir: bin
|
10
10
|
cert_chain: []
|
11
|
-
date:
|
11
|
+
date: 2022-03-13 00:00:00.000000000 Z
|
12
12
|
dependencies:
|
13
13
|
- !ruby/object:Gem::Dependency
|
14
14
|
name: libmf
|
@@ -76,7 +76,7 @@ required_rubygems_version: !ruby/object:Gem::Requirement
|
|
76
76
|
- !ruby/object:Gem::Version
|
77
77
|
version: '0'
|
78
78
|
requirements: []
|
79
|
-
rubygems_version: 3.
|
79
|
+
rubygems_version: 3.3.7
|
80
80
|
signing_key:
|
81
81
|
specification_version: 4
|
82
82
|
summary: Recommendations for Ruby and Rails using collaborative filtering
|