Improving Search Relevance in Xianyu: System Design and Model Implementation

Background – Xianyu’s search is the main traffic source for transactions. While efficiency is important, relevance and platform health are equally critical. This article focuses on relevance optimization.

Problem – Existing recall expansion strategies (query rewriting, i2i, image‑text extraction, etc.) introduce many low‑relevance cases. Analysis of top and long‑tail cases reveals insufficient coverage, inaccurate semantics, and slow data updates.

Technical Solution – Overview – The first phase aims to build a high‑coverage relevance matching pipeline and later explore finer semantic matching.

Engineering Pipeline – The classic search architecture consists of SPL (data transfer), QP (query understanding & feature extraction), Ha3 engine (recall, coarse & fine ranking), and RankService (online model re‑ranking). Relevance matching can be placed either in Ha3’s fine‑ranking plugin or RankService; the former is chosen for flexibility. Features are extracted offline and stored in KV for online retrieval.

Feature Construction – Features are divided into three groups: basic (statistics, structural matches), text‑matching (term overlap, BM25, synonym matches), and semantic matching (click‑behavior embeddings, multimodal embeddings). The article details the semantic matching approaches.

Text Semantic Matching – A dual‑tower vector model based on BERT (shared weights) is trained pointwise. Positive samples are high‑CTR items under sufficient exposure; negatives include sibling‑category items, low‑exposure items, query term replacements, and random batch negatives. The resulting model achieves >90% accuracy on sampled data and is deployed via offline embedding extraction and online similarity lookup.

Multimodal Semantic Matching – Uses a pretrained multimodal BERT with ResNet‑derived image region features (Electra‑small backbone). A triple‑loss dual‑tower is trained, improving offline AUC to 0.75, though online latency due to image processing remains a challenge.

Click‑Graph Matching – Constructs click pairs, applies the SWING algorithm to generate query‑to‑query similarity scores, and aggregates term weights to form sparse query representations used in online matching.

Feature Fusion – All relevance features are fused using a GBDT model (regression on four relevance levels). Training uses a modest manually labeled dataset (≈10k samples). The offline AUC reaches 0.86, and online experiments show +6.51% DCG@10, +24% query satisfaction, and a 20.45% increase in click‑interaction for related scenarios.

Conclusion & Future Work – The first‑phase baseline mitigates many relevance issues but leaves room for improvement in fine‑grained attribute matching, query tagging, and richer structured features. Future directions include expanding labeled data, refining feature granularity, and balancing relevance with transaction efficiency.