Task Adaptive Multi-learner Network for Joint CTR and CVR Estimation

CTR and CVR are critical factors in personalized applications, and many methods jointly estimate them via multi-task learning to alleviate the ultra-sparsity of conversion behaviors. However, it is still difficult to predict CVR accurately and robustly due to the limited and even biased knowledge extracted by the single model tower optimized on insufficient conversion samples. In this paper, we propose a task adaptive multi-learner (TAML) framework for joint CTR and CVR prediction. We design a hierarchical task adaptive knowledge representation module with different experts to capture knowledge in different granularities, which can effectively exploit the commonalities between CTR and CVR estimation tasks meanwhile keeping their unique characteristics. We apply multiple learners to extract data knowledge from various views and fuse their predictions to obtain accurate and robust scores. To facilitate knowledge sharing across learners, we further perform self-distillation that uses the fused scores to teach different learners. Thorough offline and online experiments show the superiority of TAML in different Ad ranking tasks, and we have deployed it in Huawei’s online advertising platform to serve the main traffic.