Problem
A recommender team has a hybrid model (matrix factorization plus features) and wants better quality, especially for cold users and new items. A heavy boosting model cannot score every item for every user online: with billions of records the load is too high. Two-tower networks look promising, but one-hot inputs and dense interaction matrices strain memory. The workshop compares these options on a music-service dataset.
Example from the article: a music dataset where a song is a positive if the user listened to it repeatedly and a negative if it was heard once, or two to three times.
Solution in brief
LightFM scores all user-item pairs offline, and CatBoost reranks the top candidates using the LightFM score as one more feature. Separately, DSSM-style networks map user and item features to embeddings trained with triplet loss. Every model is evaluated on a general test set and on separate cold-user and cold-item sets.
Questions this case answers
- How to improve ranking quality for new users and new items when scoring the whole catalog with a heavy model is too expensive
- How to train a two-tower retrieval model from user and item features when interaction identifiers are too sparse to use
- How to rerank a cheap first-level recommender with a gradient boosting model without scoring every item online?
- How to evaluate a recommender separately on cold users and cold items?
Tools and methods
- LightFM — first-level hybrid recommender; its score becomes the CatBoost feature.
- CatBoost — second-level reranker trained on both positives and negatives.
- Two-stage recommender — cheap candidate scoring, then reranking of the top items.
- DSSM with triplet loss — user and item networks (User2Vec, Item2Vec) trained on anchor, positive and negative.
- t-SNE and UMAP — 2D views of item embeddings colored by genre.
- SciPy sparse matrices — user and item weights and the interaction matrix.
Google or a one-time email code. No password.
