All-Rounder Recall Representation Algorithm Practice

The talk, presented by Pan Shengyi, an algorithm expert from NetEase Yanxuan, introduces the team’s focus on search and recommendation within the broader AI department that also covers NLP and supply‑chain optimization.

It begins by defining the recall representation problem: learning low‑dimensional dense embeddings for discrete IDs so that related items have high inner‑product similarity, enabling vector‑based similarity measures for ranking.

Data processing for recommendation is described as a pipeline of recall, pre‑ranking, ranking, and re‑ranking, where the recall representation model operates primarily in the recall and pre‑ranking stages, serving as a foundational component.

The capabilities of the embedding model are illustrated: a single embedding can support both recall and coarse‑ranking via vector similarity, while multiple embeddings may be combined using a coarse‑ranking model to aggregate results.

The value of such models is highlighted through three aspects: (1) broad applicability across NLP, search, recommendation, and vision; (2) mature engineering solutions, including vector search engines like Faiss and SCANN; and (3) rapid academic progress in GCN/GNN techniques that can be leveraged for further gains.

Implementation evolves iteratively. Initially the team focuses on item embeddings because items are fewer and denser than users. A session‑based embedding model, inspired by a Word2Vec‑style approach with global items, is optimized with loss‑function improvements and multi‑layer side‑information embeddings.

Graph‑based models are then introduced. The GCN framework is explained in three stages—message passing, aggregation, and update—followed by practical variants such as GraphSAGE (sampling‑based), LightGCN (compressed matrix operations), and a custom GCN adaptation for large‑scale industrial data.

User embeddings are obtained via two strategies: a fast heuristic that aggregates item vectors with time decay, frequency weighting, and attention; and a classic DNN (YouTube‑style) that learns user vectors through supervised next‑item prediction. Session‑level representation is also explored using graph‑based session models and multi‑interest techniques like clustering‑based vectors and the MIND capsule‑network approach.

Extensive offline experiments compare models using HitRate and NDCG. Baselines include a single‑vector session model, while advanced models such as GraphSAGE, LightGCN, YouTube‑DNN, SR‑GNN, multi‑vector clustering, and MIND show varying improvements, with SR‑GNN achieving the best overall performance.

For production, the team builds a unified recall service that vectorizes users and items, supports both batch (T+1) and real‑time data, and feeds downstream ranking and re‑ranking modules for search, recommendation, and advertising, dramatically improving conversion and GMV.

Real‑world impact examples include vector‑based U2I, I2I, and Q2I retrieval, personalized banner generation, and other downstream applications that benefit from the “everything is a vector” paradigm.

The presentation concludes with acknowledgments and invitations to join the DataFunTalk community for further AI and big‑data discussions.