#2297 [FEATURE] Two-Stage Recommendation Pipeline: Candidate Generation + LightGBM Ranker

Description

Motivation

Two-stage recommendation systems are widely adopted in industry (YouTube, Pinterest, Alibaba, etc.), yet there is no end-to-end, well-structured reference implementation in this repository that clearly demonstrates how to properly connect the candidate generation stage to the ranking stage.

Most existing examples either:

Focus on a single model in isolation, or
Skip over the practical details of how to wire the two stages together correctly (feature leakage, negative sampling strategy, group construction for LTR, etc.)

This makes it difficult for practitioners to apply two-stage modeling correctly in real-world pipelines. The goal of this contribution is to provide a clean, reusable reference that fills that gap.

Proposed Design

Stage 1 — Candidate Generation

Retrieve a small set of candidate items (e.g. top-K) from the full item catalog for each user.

Status: open to suggestions. I am still evaluating which model to use for this stage. Possible directions include:

Matrix Factorization (ALS, SVD) — fast, well-understood baseline

Neural Collaborative Filtering (NCF) — already in the repo, easy to integrate

Two-Tower / DSSM — more scalable for large catalogs

SAR — lightweight, already supported

If you have a recommendation or strong opinion on which candidate model fits best here, please leave a comment.

Stage 2 — Ranking

Re-rank the retrieved candidates using LightGBM with the lambdarank objective (lgb.LGBMRanker), which was recently merged into staging via #2288.

Why This Matters

Two-stage ranking is the dominant paradigm in production recommendation systems, but getting the details right is non-trivial:

Candidate recall must be high enough that the ranker has good items to work with
Negative sampling must be done carefully to avoid inflated metrics
Feature computation must respect train/test boundaries to prevent leakage

#2288 — LightGBM LambdaRank notebook (staging, merged)
#2205 — Embedding ranker in PyTorch

Expected behavior with the suggested feature

Select and implement candidate generation model (Stage 1)
Build a connector that converts Stage 1 output into a ranked candidate pool for Stage 2
Apply proper negative sampling for the ranking stage
Evaluate end-to-end with standard metrics: NDCG@K, Precision@K, Recall@K
Provide a clean notebook (lightgbm_two_stage_movielens.ipynb) on MovieLens 1M

Willingness to contribute

Yes, I can contribute for this issue independently.
Yes, I can contribute for this issue with guidance from Recommenders community.
No, I cannot contribute at this time.

[FEATURE] Two-Stage Recommendation Pipeline: Candidate Generation + LightGBM Ranker