Hello everyone! My name is Sasha, I am CTO & Co-Founder at LoyaltyLab. Two years ago, I went with friends, like all poor students, in the evening for a beer to the nearest convenience store. We were very upset that the retailer, knowing that we would come for a beer, did not offer a discount on chips or crackers, although this is so logical! We did not understand why this situation was happening and decided to make our company. Well, as a bonus, write yourself discounts every Friday for the very chips.

And it all went so far that I speak at NVIDIA GTC with material on the technical side of the product . We are happy to share our best practices with the community, so I am posting my report in the form of an article.

Introduction

Like everything at the beginning of the journey, we started with a review of how recommendation systems are made. And the most popular was the architecture of the following type:

:

1. , .
2. c .

:

• candidate/ — user-product, .
• candidates extraction/extractor/ — “ ” .

. — ALS. , , . , , . , ALS, baseline’.

, , , realtime , , 30 , , , . , , , , — , .

baseline , ALS . , , :

• Precision — .
• Recall — , target .
• F1-score — F-, .

. 3 :

• precision@5 — -5 .
• response-rate@5 — ( 5 ).
• avg roc-auc per user — roc-auc .

, time-series -, k , test k+1 . / . -, — . , .

, , baseline, ALS .
:

:

, . , , , , “ , ”. , , :

1. .
2. , .
3. .
4. /.
5. ( ).
6. , , (Word2Vec, DSSM, etc.).

, . , K ( 1-3 ), K . , :

, , , recall precision . “ 2 ”.

,

, “ ” , , , - . . , ALS, , . “ ” , , / , , , , , . 2 :

k — , 14 .
, , .

, - . , //. , - . — , k . , . :

, , , . :

2 , :

1. , , K .
2. , .

, :

ft — , K ( K=5)
tm — ,

, (0, 0) recall precision, . , , k , .

— , . -k “” / “” /. “” “” . . , , , . . :

“category” .

, , , “” .

, — . 2 , [ i]-[ j], j > i, . product_i-product_j , product_i . -k , . :

:

, : , , . , , , .

:

k — , .
, k=4. 4 . 

, ,

— , . 3:

• ALS
• Word2Vec (Item2Vec )
• DSSM

ALS , , . Word2Vec, gensim. , , . , , “” ( ). ecommerce , Ozon. DSSM . Microsoft, , research paper. :

Q — query, , D[i] — document, -. , . (multilayer perceptron). , .
, , — . :

— ALS DSSM , Word2Vec . 3 :

1. , , .
2. .
3. TF-IDF .

, , , , . 1, , ½, ⅓, .:

TF-IDF , TF-IDF , , , , — . , . :

. ALS:

Item2Vec :

, baseline. , k . , , 50-70 .

DSSM:

?

, , ? ? :

1. - . , , , .
2. ?

, : , . , , , “ ” “ ”.

, , : (recall). , , . .

— , recall — . 2 , :

• , .
• .

, «, » , . , “ ” . , . , .

:

:

, , . , , , . , . , , :

1. / .
2. .
3. .

, , .  , . . , :

1. / .
2. , . , -.
3. , , .

:

overall products conversion — , .

. , , , . , , .

, , , , — 3-4%:

, , realtime , . , .

, . telegram. AI/ telegram — welcome :)