Exploring Generalized Multi‑Objective Recommendation Algorithms for 58 Community

Guest speaker Zhou Jianbin, senior algorithm architect at 58.com, shares the business background of 58 Community, a content feed platform serving local users with PGC and UGC.

The platform aims to increase the proportion of value‑content (real‑estate, cars, jobs, local life) while maintaining click‑through rate, retention and interaction metrics.

Initially the goal was single‑objective click‑rate optimization; later it evolved to multi‑objective targets such as value‑content share, stable click‑rate, and user retention, requiring multi‑objective ranking models (e.g., shared‑bottom ESSM, MM‑OE).

To boost value‑content share, the team moved from rule‑based boosting to improving the ranking model by incorporating cross‑domain user behavior from other 58 services, treating them as auxiliary features.

The first attempt modified the DeepFM+DIN model by simplifying DIN features and replacing concatenation with pooling for heterogeneous embeddings, but suffered from slow training and negligible impact due to sparse and mismatched cross‑domain items.

They then adopted a cross‑domain embedding pre‑training approach inspired by Alibaba’s EGES, extracting core attributes from each business line, compressing IDs, and constructing a weighted user‑behavior graph for DeepWalk‑style random walks, yielding joint item and attribute embeddings.

Replacing the original DIN embeddings with EGES‑pre‑trained vectors improved both click‑through rate and value‑content proportion (from 12% to 28%).

For the long‑term retention goal, they identified four key features: interaction rate, first‑visit content weight, last‑visit content weight, and diversity. A re‑ranking formula combines click‑prediction with weighted contributions of these features, and a diversity adjustment factor θ is applied.

Parameter optimization is performed online using the Cross‑Entropy Method (CEM), sampling parameter sets, evaluating rewards (weighted retention, click‑rate, interaction), selecting top‑k, and iterating, allowing continuous adaptation without exhaustive A/B tests.

After about ten CEM iterations, next‑day retention improved by 1% while click‑rate and interaction remained stable, confirming the effectiveness of the multi‑objective and online optimization framework.

Future work includes exploring reinforcement learning for multi‑objective recommendation to further enhance long‑term ecosystem health.