Personalized Recommendation for Xianyu Small Item Pools: Challenges and Solutions

Introduction Personalization is an effective way to improve conversion rates. Xianyu has applied personalized recommendation algorithms in large-scale scenarios, but the emergence of micro, fine‑grained item pools (under ten thousand items) poses new challenges due to short item lifecycles and high real‑time requirements.

1. Xianyu Recommendation Workflow The system consists of a Match (recall) stage and a Rank (scoring) stage. User triggers (clicks, purchases, likes, etc.) are combined with an X2I matrix to retrieve candidate items, which are then scored and sorted. Figure 1 illustrates the eight‑step pipeline, including user trigger extraction, item recall via X2I, filtering, scoring, ranking, and truncation.

2. Characteristics of Xianyu Small Item Pools Items are unique, have short lifespans, and possess subjective value attributes. The pools are highly real‑time and often periodic (e.g., 6‑hour auction cycles), requiring fresh item inflow and rapid turnover.

3. Problems in Applying the Main‑Site Flow Experiments on pools ranging from 20 k to 1 M items show severe recall deficiency: for a 200 k pool with 1 000 recall requests, only 3 items meet the rules. The main causes are (1) reliance on a global X2I matrix that yields few matches after pool filtering, and (2) hour‑level X2I refresh that cannot keep up with sub‑hourly pool dynamics.

4. Proposed Solutions

4.1 Pre‑filtering Filter items before they are fed back to the search engine, generating a pool‑specific X2I to improve recall.

4.2 Vector Search Encode items and user triggers into vectors (≈200‑dimensional) and perform similarity search. This guarantees recall and supports minute‑level freshness.

4.3 Search‑Engine Recall Use a real‑time search engine with item tags and rules. User triggers become query terms, and the engine returns candidates with built‑in timeliness.

5. Implementation Details Items are structured into a wide table, enriched with quality scores, tag predictions, and rule information, then synced to the search engine (full and incremental). At query time, the request is expanded tenfold, results are de‑duplicated per category (max 3 per category), filtered against a 200‑length FIFO of previously viewed items, and finally ranked and truncated. The pipeline is illustrated in Figure 7.

6. Online Results A bucket‑test shows that after deploying the search‑engine based recall, per‑user transaction count increased by 0.14 %, clicks rose by 14.1 %, and click‑user count grew by 18.2 %.

Conclusion The standard Xianyu recommendation pipeline suffers from recall shortage in small item pools. Among three explored remedies, the search‑engine based approach proved most effective and was successfully deployed, yielding measurable improvements in engagement and conversion.