Multi-Channel Deep Interest Modeling for 58.com Home Page Recommendations

The 58.com platform, with hundreds of millions of daily active users across housing, recruitment, second‑hand goods, and more, faces complex multi‑business recommendation challenges on its homepage. To balance connection efficiency, revenue, user experience, and operational goals, a deep learning model evolving from a single‑channel to a dual‑channel architecture was implemented.

Challenges of Multi‑Business Fusion include adapting recall strategies to multiple channels, ensuring balanced traffic distribution in re‑ranking, handling diverse display styles, and overcoming feature alignment and engineering complexity for traditional XGBoost ranking models.

Behavior Sequence Interest Model Validation shows that sequence‑based models (DIN, DIEN, Transformer) improve interest modeling while reducing feature engineering effort, but pure sequence models still lag behind feature‑rich XGBoost. Therefore, high‑level business features were incorporated into the deep model.

Customized Channel Introduction embeds feature‑crossing logic into the model via four layers: feature input, vectorization (pre‑trained embeddings), cross‑layer (e.g., cosine similarity, DNN), and concatenation, enabling business‑specific feature interactions.

Model Architecture combines a customized channel and a serialized behavior channel, both feeding into an MLP, forming the dual‑channel model that improves click‑through rate by 3% and exposure conversion by 5% while simplifying feature engineering.

Multi‑Channel Model Exploration extends the architecture with additional behavior channels (click, search, conversion, content) to capture comprehensive user interests, achieving over 10% lift in exposure conversion and supporting scenario‑specific adaptations.

Engineering Practices include reducing training time from five days to five hours through data format redesign, optimizing online latency from 10% timeout to 0.3% by batch sizing and vector layer offloading to Redis, and aligning offline‑online behavior sequences by shifting training windows two minutes earlier.

Future Directions focus on multi‑objective optimization (e.g., reinforcement learning) and further multi‑channel enhancements such as negative feedback modeling and window‑based collaborative representations.

Overall, the multi‑channel deep interest modeling framework significantly boosts recommendation performance across 58.com’s diverse services.