Vectorized Recall and Dual‑Tower Model for Home Page Recommendation at 58.com

Background 58.com operates a massive lifestyle information platform; its home‑page “You May Like” recommendation serves millions of users and billions of items, requiring efficient and diverse recall mechanisms.

Vectorized Recall with Word2Vec The team first added a vector‑based recall channel using Word2Vec embeddings trained on user click sequences. Two training modes (CBOW and Skip‑gram) were evaluated, with Skip‑gram chosen for better representation of low‑frequency items. Data preprocessing involved extracting ItemID, UserID, timestamps from HDFS logs, constructing session‑aware sequences, and applying city‑ and business‑aware negative sampling to improve relevance.

Optimizations included limiting negative samples to the same city, adjusting sampling ranges, and adding business constraints, which increased training time but yielded better offline and online performance. Online prediction aggregates user click vectors, averages them to form a user vector, and retrieves the top‑50 similar items via a FAISS inverted‑index cluster, reducing latency from 35 ms to 10 ms.

Dual‑Tower Model Deployment To overcome limitations of Word2Vec (uniform user‑vector weighting and lack of contextual features), a dual‑tower architecture was introduced. The model consists of an input layer (user tower and item tower), a representation layer (embedding of ItemID, LocalID, CateID, and pre‑trained Word2Vec vectors), and a matching layer (cosine similarity and softmax). Negative sampling combines easy random samples with hard samples from the same city but different business categories.

Additional user features such as long‑term interest tags, device attributes, and behavior statistics are incorporated via attention mechanisms, enhancing the user embedding. Offline evaluation uses AUC, coverage, and average position metrics, showing consistent improvements when adding interest tags and user features.

Online serving mirrors the offline pipeline: after training, item embeddings are generated and stored; at request time, user features produce a user embedding, which is used to query the FAISS index for top‑50 items. Multi‑category parallel recall further refines results by separating behavior sequences per business line before averaging.

Summary and Outlook Vectorized recall significantly boosts click‑through and conversion rates. Future work includes finer‑grained user interest modeling, richer item attributes (price, layout), advanced hard‑negative sampling strategies, and potential boosting‑style training to further improve recommendation quality.