Yandex Programming Championship 2019

Machine Learining track

Stage 2

Recommender system

Given historical user's transaction develop a recommender system for online-gallery (description of the task in russian).

Appoach

Take all transactions (clicks, likes and bookmarks) and create rating matrix with value

$$r_0 + s^{md - d} + esp + 1$$,

with initial rating $$r_0=50$$, smoothing parameter $$s=0.99$$, maximum transaction date $$md$$, current date of transaction $$d$$ and small tolerance $$e=10^{-5}$$. Train 2 models from implicit:

BM25Recommender with 600 neighbours,
AlternatingLeastSquares with 512 factors and 15 iterations

and predict top 400 candidates with each model. Then blend these 3 models with weights 5.5 and 6 for BM25Recommender and AlternatingLeastSquares respectively. The blending consist in rearranging of candidates (see mix_solutions in train.py).

This will reach a mean average precision at 100 (mAP@100) multiplied by 10000 around 29 (see leaderboard).

Usage