Building and Applying a Vector System for Search and Recommendation at NetEase Yanxuan

Since the second half of 2018, NetEase Yanxuan has been exploring the use of vector representations in its search and recommendation pipelines, gradually expanding from product recall to search ranking, discovery, suggestion, cross‑category recommendation, multi‑interest recall, and real‑time re‑ranking.

The core idea is to embed all relevant entities—users (U), items (I), queries (Q), categories (C), topics (T), etc.—into a shared vector space, enabling similarity calculations for a wide range of matching tasks such as U→I, Q→I, I→I, and beyond.

Initially, graph embedding techniques like LINE and Node2Vec were tried, followed by a YouTubeDNN‑style approach. Ultimately a two‑step strategy was adopted: first learning high‑quality item vectors, then aggregating other entities through a dedicated aggregation module that incorporates interaction frequency, time decay, noise filtering, attention mechanisms, and optional clustering to produce multi‑vector representations.

Item vector learning relies on rich user‑behavior signals (clicks, purchases, co‑purchases) and item attributes, combining them in a loss that jointly optimizes center‑context and center‑global similarities with negative sampling. Incremental training and affine‑transform alignment are used to keep daily vectors in a consistent space.

Beyond items, the system incrementally learns vectors for users, queries, categories, and topics, allowing flexible U→I, Q→I, Q→C, etc., matching. A unified aggregation framework makes it easy to add new entity types by supplying a target‑source relationship table.

To handle billions of vectors, offline block‑matrix similarity computation and online approximate nearest‑neighbor search (LSH, FAISS) are employed, supporting both batch and real‑time retrieval.

In search scenarios, vector similarity powers query suggestion, discovery words, and personalized ranking by combining long‑term, short‑term, and real‑time user vectors with item vectors, improving relevance and conversion rates.

For recommendation recall, multiple user vectors (long‑term, short‑term, real‑time, multi‑interest, group) are used to balance precision and diversity across different placement slots.

Purchase prediction leverages I→I similarity to distinguish focused browsing from casual wandering, further refined with historical user‑item interactions (U→I) to boost conversion.

The system demonstrates the practical advantages of a unified vector approach—rapid feature rollout, superior performance over traditional methods, and a clear path for future extensions beyond vector embeddings.