Effective Knowledge Transfer for Multi-Task Recommendation Models

The conversion rate (CVR) is critical for platform effectiveness, but sparse customer conversion actions make training ranking models challenging. In the new paper "Effective Knowledge Transfer for Multi-task Recommendation Models," authors Guohao Cai, Jun Yuan, and Zhenhua Dong introduce EKTM to address this. The method enables a ranking model to learn from diverse user behaviors by transferring knowledge across related tasks. A router module collects and disseminates knowledge across tasks, while each CVR task includes a transmitter module that transforms the router's knowledge into task-specific insights. An additional enhancement module ensures transferred knowledge benefits original task learning without interference.

Extensive experiments on several benchmark datasets show EKTM outperforms existing state-of-the-art multi-task recommendation approaches. The real-world impact was validated through online A/B testing on a commercial platform, where EKTM achieved a 3.93% uplift in effective Cost Per Mille (eCPM). The algorithm has been fully deployed across two of the platform's main-traffic scenarios, demonstrating its scalability and effectiveness in large-scale industrial settings. The paper is available on arXiv as cs.IR/2605.05730.