disco 0.2.4 → 0.2.5
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +4 -4
- data/CHANGELOG.md +8 -0
- data/README.md +23 -5
- data/lib/disco.rb +1 -0
- data/lib/disco/metrics.rb +10 -0
- data/lib/disco/recommender.rb +154 -74
- data/lib/disco/version.rb +1 -1
- metadata +3 -2
checksums.yaml
CHANGED
@@ -1,7 +1,7 @@
|
|
1
1
|
---
|
2
2
|
SHA256:
|
3
|
-
metadata.gz:
|
4
|
-
data.tar.gz:
|
3
|
+
metadata.gz: 8fbecb858b316ed39a9cb726263e182561cba6df498e6253d88c79ebec5cab05
|
4
|
+
data.tar.gz: 42eb38a6e4e0b3fc5a9452deae5a48676ae9a53e78eeb6197718a0c94bd02b6b
|
5
5
|
SHA512:
|
6
|
-
metadata.gz:
|
7
|
-
data.tar.gz:
|
6
|
+
metadata.gz: d0250346d75fba75064a29578f6bfd39f09ecf712ba2e505b97a4952b5ff8b31af307eb1b912e9b25cc3dc28dee0d096bea44b47bb2ef268859bb4171f0ef8b2
|
7
|
+
data.tar.gz: 7b341328c12885efd0ffece4201036bb9457caee80a48a99ba110af9a81bcf832bbc1e8f8f5f14e7fddffef2dd3f4643837e0d569c997ab0c2d9ae85e12422f7
|
data/CHANGELOG.md
CHANGED
@@ -1,3 +1,11 @@
|
|
1
|
+
## 0.2.5 (2021-02-20)
|
2
|
+
|
3
|
+
- Added `top_items` method
|
4
|
+
- Added `optimize_similar_users` method
|
5
|
+
- Added support for Faiss for `optimize_item_recs` and `optimize_similar_users` methods
|
6
|
+
- Added `rmse` method
|
7
|
+
- Improved performance
|
8
|
+
|
1
9
|
## 0.2.4 (2021-02-15)
|
2
10
|
|
3
11
|
- Added `user_ids` and `item_ids` methods
|
data/README.md
CHANGED
@@ -201,6 +201,8 @@ bin = File.binread("recommender.bin")
|
|
201
201
|
recommender = Marshal.load(bin)
|
202
202
|
```
|
203
203
|
|
204
|
+
Alternatively, you can store only the factors and use a library like [Neighbor](https://github.com/ankane/neighbor)
|
205
|
+
|
204
206
|
## Algorithms
|
205
207
|
|
206
208
|
Disco uses high-performance matrix factorization.
|
@@ -237,6 +239,16 @@ There are a number of ways to deal with this, but here are some common ones:
|
|
237
239
|
- For user-based recommendations, show new users the most popular items.
|
238
240
|
- For item-based recommendations, make content-based recommendations with a gem like [tf-idf-similarity](https://github.com/jpmckinney/tf-idf-similarity).
|
239
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) your application’s Gemfile) and item frequency for implicit feedback.
|
251
|
+
|
240
252
|
## Data
|
241
253
|
|
242
254
|
Data can be an array of hashes
|
@@ -257,23 +269,29 @@ Or a Daru data frame
|
|
257
269
|
Daru::DataFrame.from_csv("ratings.csv")
|
258
270
|
```
|
259
271
|
|
260
|
-
##
|
272
|
+
## Performance [master]
|
261
273
|
|
262
|
-
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.
|
263
275
|
|
264
276
|
Add this line to your application’s Gemfile:
|
265
277
|
|
266
278
|
```ruby
|
267
|
-
gem '
|
279
|
+
gem 'faiss'
|
280
|
+
```
|
281
|
+
|
282
|
+
Speed up the `user_recs` method with:
|
283
|
+
|
284
|
+
```ruby
|
285
|
+
model.optimize_user_recs
|
268
286
|
```
|
269
287
|
|
270
|
-
Speed up
|
288
|
+
Speed up the `item_recs` method with:
|
271
289
|
|
272
290
|
```ruby
|
273
291
|
model.optimize_item_recs
|
274
292
|
```
|
275
293
|
|
276
|
-
Speed up
|
294
|
+
Speed up the `similar_users` method with:
|
277
295
|
|
278
296
|
```ruby
|
279
297
|
model.optimize_similar_users
|
data/lib/disco.rb
CHANGED
data/lib/disco/recommender.rb
CHANGED
@@ -2,12 +2,13 @@ module Disco
|
|
2
2
|
class Recommender
|
3
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
9
|
@user_map = {}
|
10
10
|
@item_map = {}
|
11
|
+
@top_items = top_items
|
11
12
|
end
|
12
13
|
|
13
14
|
def fit(train_set, validation_set: nil)
|
@@ -41,6 +42,16 @@ module Disco
|
|
41
42
|
end
|
42
43
|
@rated.default = nil
|
43
44
|
|
45
|
+
if @top_items
|
46
|
+
@item_count = [0] * @item_map.size
|
47
|
+
@item_sum = [0.0] * @item_map.size
|
48
|
+
train_set.each do |v|
|
49
|
+
i = @item_map[v[:item_id]]
|
50
|
+
@item_count[i] += 1
|
51
|
+
@item_sum[i] += (v[value_key] || 1)
|
52
|
+
end
|
53
|
+
end
|
54
|
+
|
44
55
|
eval_set = nil
|
45
56
|
if validation_set
|
46
57
|
eval_set = []
|
@@ -67,8 +78,9 @@ module Disco
|
|
67
78
|
@user_factors = model.p_factors(format: :numo)
|
68
79
|
@item_factors = model.q_factors(format: :numo)
|
69
80
|
|
70
|
-
@
|
71
|
-
@
|
81
|
+
@user_recs_index = nil
|
82
|
+
@similar_users_index = nil
|
83
|
+
@similar_items_index = nil
|
72
84
|
end
|
73
85
|
|
74
86
|
# generates a prediction even if a user has already rated the item
|
@@ -95,61 +107,76 @@ module Disco
|
|
95
107
|
u = @user_map[user_id]
|
96
108
|
|
97
109
|
if u
|
98
|
-
|
99
|
-
|
100
|
-
predictions =
|
101
|
-
@item_map.keys.zip(predictions).map do |item_id, pred|
|
102
|
-
{item_id: item_id, score: pred}
|
103
|
-
end
|
110
|
+
rated = item_ids ? {} : @rated[u]
|
104
111
|
|
105
112
|
if item_ids
|
106
|
-
|
107
|
-
|
113
|
+
ids = Numo::NArray.cast(item_ids.map { |i| @item_map[i] }.compact)
|
114
|
+
return [] if ids.size == 0
|
115
|
+
|
116
|
+
predictions = @item_factors[ids, true].inner(@user_factors[u, true])
|
117
|
+
indexes = predictions.sort_index.reverse
|
118
|
+
indexes = indexes[0...[count + rated.size, indexes.size].min] if count
|
119
|
+
predictions = predictions[indexes]
|
120
|
+
ids = ids[indexes]
|
121
|
+
elsif @user_recs_index && count
|
122
|
+
predictions, ids = @user_recs_index.search(@user_factors[u, true].expand_dims(0), count + rated.size).map { |v| v[0, true] }
|
108
123
|
else
|
109
|
-
@
|
110
|
-
|
111
|
-
|
124
|
+
predictions = @item_factors.inner(@user_factors[u, true])
|
125
|
+
# TODO make sure reverse isn't hurting performance
|
126
|
+
indexes = predictions.sort_index.reverse
|
127
|
+
indexes = indexes[0...[count + rated.size, indexes.size].min] if count
|
128
|
+
predictions = predictions[indexes]
|
129
|
+
ids = indexes
|
112
130
|
end
|
113
131
|
|
114
|
-
predictions.
|
115
|
-
predictions = predictions.first(count) if count && !item_ids
|
132
|
+
predictions.inplace.clip(@min_rating, @max_rating) if @min_rating
|
116
133
|
|
117
|
-
|
118
|
-
|
119
|
-
|
120
|
-
|
121
|
-
pred[:score] = pred[:score].clamp(@min_rating, @max_rating)
|
122
|
-
end
|
123
|
-
end
|
134
|
+
keys = @item_map.keys
|
135
|
+
result = []
|
136
|
+
ids.each_with_index do |item_id, i|
|
137
|
+
next if rated[item_id]
|
124
138
|
|
125
|
-
|
139
|
+
result << {item_id: keys[item_id], score: predictions[i]}
|
140
|
+
break if result.size == count
|
141
|
+
end
|
142
|
+
result
|
143
|
+
elsif @top_items
|
144
|
+
top_items(count: count)
|
126
145
|
else
|
127
|
-
# no items if user is unknown
|
128
|
-
# TODO maybe most popular items
|
129
146
|
[]
|
130
147
|
end
|
131
148
|
end
|
132
149
|
|
133
|
-
def
|
150
|
+
def similar_items(item_id, count: 5)
|
134
151
|
check_fit
|
135
|
-
@
|
152
|
+
similar(item_id, @item_map, item_norms, count, @similar_items_index)
|
136
153
|
end
|
137
|
-
alias_method :
|
154
|
+
alias_method :item_recs, :similar_items
|
138
155
|
|
139
|
-
def
|
156
|
+
def similar_users(user_id, count: 5)
|
140
157
|
check_fit
|
141
|
-
@
|
158
|
+
similar(user_id, @user_map, user_norms, count, @similar_users_index)
|
142
159
|
end
|
143
160
|
|
144
|
-
def
|
161
|
+
def top_items(count: 5)
|
145
162
|
check_fit
|
146
|
-
|
147
|
-
end
|
148
|
-
alias_method :item_recs, :similar_items
|
163
|
+
raise "top_items not computed" unless @top_items
|
149
164
|
|
150
|
-
|
151
|
-
|
152
|
-
|
165
|
+
if @implicit
|
166
|
+
scores = @item_count
|
167
|
+
else
|
168
|
+
require "wilson_score"
|
169
|
+
|
170
|
+
range = @min_rating..@max_rating
|
171
|
+
scores = @item_sum.zip(@item_count).map { |s, c| WilsonScore.rating_lower_bound(s / c, c, range) }
|
172
|
+
end
|
173
|
+
|
174
|
+
scores = scores.map.with_index.sort_by { |s, _| -s }
|
175
|
+
scores = scores.first(count) if count
|
176
|
+
item_ids = item_ids()
|
177
|
+
scores.map do |s, i|
|
178
|
+
{item_id: item_ids[i], score: s}
|
179
|
+
end
|
153
180
|
end
|
154
181
|
|
155
182
|
def user_ids
|
@@ -178,17 +205,61 @@ module Disco
|
|
178
205
|
end
|
179
206
|
end
|
180
207
|
|
208
|
+
def optimize_user_recs
|
209
|
+
check_fit
|
210
|
+
@user_recs_index = create_index(item_factors, library: "faiss")
|
211
|
+
end
|
212
|
+
|
213
|
+
def optimize_similar_items(library: nil)
|
214
|
+
check_fit
|
215
|
+
@similar_items_index = create_index(item_norms, library: library)
|
216
|
+
end
|
217
|
+
alias_method :optimize_item_recs, :optimize_similar_items
|
218
|
+
|
219
|
+
def optimize_similar_users(library: nil)
|
220
|
+
check_fit
|
221
|
+
@similar_users_index = create_index(user_norms, library: library)
|
222
|
+
end
|
223
|
+
|
181
224
|
private
|
182
225
|
|
183
|
-
|
184
|
-
|
226
|
+
# factors should already be normalized for similar users/items
|
227
|
+
def create_index(factors, library:)
|
228
|
+
# TODO make Faiss the default in 0.3.0
|
229
|
+
library ||= defined?(Faiss) && !defined?(Ngt) ? "faiss" : "ngt"
|
230
|
+
|
231
|
+
case library
|
232
|
+
when "faiss"
|
233
|
+
require "faiss"
|
234
|
+
|
235
|
+
# inner product is cosine similarity with normalized vectors
|
236
|
+
# https://github.com/facebookresearch/faiss/issues/95
|
237
|
+
#
|
238
|
+
# TODO use non-exact index
|
239
|
+
# https://github.com/facebookresearch/faiss/wiki/Faiss-indexes
|
240
|
+
index = Faiss::IndexFlatIP.new(factors.shape[1])
|
241
|
+
|
242
|
+
# ids are from 0...total
|
243
|
+
# https://github.com/facebookresearch/faiss/blob/96b740abedffc8f67389f29c2a180913941534c6/faiss/Index.h#L89
|
244
|
+
index.add(factors)
|
245
|
+
|
246
|
+
index
|
247
|
+
when "ngt"
|
248
|
+
require "ngt"
|
249
|
+
|
250
|
+
# could speed up search with normalized cosine
|
251
|
+
# https://github.com/yahoojapan/NGT/issues/36
|
252
|
+
index = Ngt::Index.new(factors.shape[1], distance_type: "Cosine")
|
185
253
|
|
186
|
-
|
187
|
-
|
188
|
-
|
189
|
-
|
190
|
-
|
191
|
-
|
254
|
+
# NGT normalizes so could call create_index with factors instead of norms
|
255
|
+
# but keep code simple for now
|
256
|
+
ids = index.batch_insert(factors)
|
257
|
+
raise "Unexpected ids. Please report a bug." if ids.first != 1 || ids.last != factors.shape[0]
|
258
|
+
|
259
|
+
index
|
260
|
+
else
|
261
|
+
raise ArgumentError, "Invalid library: #{library}"
|
262
|
+
end
|
192
263
|
end
|
193
264
|
|
194
265
|
def user_norms
|
@@ -202,40 +273,38 @@ module Disco
|
|
202
273
|
def norms(factors)
|
203
274
|
norms = Numo::SFloat::Math.sqrt((factors * factors).sum(axis: 1))
|
204
275
|
norms[norms.eq(0)] = 1e-10 # no zeros
|
205
|
-
norms
|
276
|
+
factors / norms.expand_dims(1)
|
206
277
|
end
|
207
278
|
|
208
|
-
def similar(id, map,
|
279
|
+
def similar(id, map, norm_factors, count, index)
|
209
280
|
i = map[id]
|
210
|
-
|
281
|
+
|
282
|
+
if i && norm_factors.shape[0] > 1
|
211
283
|
if index && count
|
212
|
-
|
213
|
-
|
214
|
-
|
215
|
-
|
216
|
-
|
217
|
-
|
218
|
-
|
219
|
-
|
220
|
-
}
|
284
|
+
if defined?(Faiss) && index.is_a?(Faiss::Index)
|
285
|
+
predictions, ids = index.search(norm_factors[i, true].expand_dims(0), count + 1).map { |v| v.to_a[0] }
|
286
|
+
else
|
287
|
+
result = index.search(norm_factors[i, true], size: count + 1)
|
288
|
+
# ids from batch_insert start at 1 instead of 0
|
289
|
+
ids = result.map { |v| v[:id] - 1 }
|
290
|
+
# convert cosine distance to cosine similarity
|
291
|
+
predictions = result.map { |v| 1 - v[:distance] }
|
221
292
|
end
|
222
293
|
else
|
223
|
-
|
224
|
-
|
225
|
-
|
226
|
-
|
227
|
-
|
228
|
-
|
229
|
-
|
230
|
-
|
231
|
-
|
232
|
-
|
233
|
-
|
234
|
-
|
235
|
-
|
236
|
-
|
237
|
-
predictions.each { |pred| pred[:score] /= norms[i] }
|
238
|
-
predictions
|
294
|
+
predictions = norm_factors.inner(norm_factors[i, true])
|
295
|
+
indexes = predictions.sort_index.reverse
|
296
|
+
indexes = indexes[0...[count + 1, indexes.size].min] if count
|
297
|
+
predictions = predictions[indexes]
|
298
|
+
ids = indexes
|
299
|
+
end
|
300
|
+
|
301
|
+
keys = map.keys
|
302
|
+
|
303
|
+
# TODO use user_id for similar_users in 0.3.0
|
304
|
+
key = :item_id
|
305
|
+
|
306
|
+
(1...ids.size).map do |i|
|
307
|
+
{key => keys[ids[i]], score: predictions[i]}
|
239
308
|
end
|
240
309
|
else
|
241
310
|
[]
|
@@ -304,6 +373,11 @@ module Disco
|
|
304
373
|
obj[:max_rating] = @max_rating
|
305
374
|
end
|
306
375
|
|
376
|
+
if @top_items
|
377
|
+
obj[:item_count] = @item_count
|
378
|
+
obj[:item_sum] = @item_sum
|
379
|
+
end
|
380
|
+
|
307
381
|
obj
|
308
382
|
end
|
309
383
|
|
@@ -320,6 +394,12 @@ module Disco
|
|
320
394
|
@min_rating = obj[:min_rating]
|
321
395
|
@max_rating = obj[:max_rating]
|
322
396
|
end
|
397
|
+
|
398
|
+
@top_items = obj.key?(:item_count)
|
399
|
+
if @top_items
|
400
|
+
@item_count = obj[:item_count]
|
401
|
+
@item_sum = obj[:item_sum]
|
402
|
+
end
|
323
403
|
end
|
324
404
|
end
|
325
405
|
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: 0.2.5
|
5
5
|
platform: ruby
|
6
6
|
authors:
|
7
7
|
- Andrew Kane
|
8
8
|
autorequire:
|
9
9
|
bindir: bin
|
10
10
|
cert_chain: []
|
11
|
-
date: 2021-02-
|
11
|
+
date: 2021-02-20 00:00:00.000000000 Z
|
12
12
|
dependencies:
|
13
13
|
- !ruby/object:Gem::Dependency
|
14
14
|
name: libmf
|
@@ -51,6 +51,7 @@ files:
|
|
51
51
|
- lib/disco.rb
|
52
52
|
- lib/disco/data.rb
|
53
53
|
- lib/disco/engine.rb
|
54
|
+
- lib/disco/metrics.rb
|
54
55
|
- lib/disco/model.rb
|
55
56
|
- lib/disco/recommender.rb
|
56
57
|
- lib/disco/version.rb
|