Click-aware Structure Transfer with Sample Weight Assignment (CSTWA) for Multi‑task CVR Optimization

Motivation The paper addresses two key issues in CVR learning: (1) the overwhelming amount of CTR knowledge can dominate CVR training, and (2) conflicts between CTR and CVR signals (e.g., clicks without purchases) lead to contradictory gradients, a phenomenon termed the "curse of knowledge".

Method CSTWA consists of three components: (1) Structure Migrator extracts high‑order CTR information via a similarity graph and transfers it to CVR; (2) Click Perceptron learns a bias vector from CTR outputs and calibrates CVR features through element‑wise multiplication and addition; (3) Curse Escaper detects label‑level conflicts between high CTR/low CVR and low CTR/high CVR cases, adjusting sample weights to weaken harmful gradients and strengthen useful ones.

Main Results Experiments on Shopee's industrial data and a public Alibaba dataset show that CSTWA outperforms single‑task baselines and other multi‑task models (MMoE, PLE, ESMM, AITM). The method achieves the highest CVR AUC, while CTR performance may not improve, reflecting the focus on CVR quality.

Conclusions CSTWA effectively mitigates the curse of knowledge by enhancing knowledge sharing and reducing conflict through its three modules. Extensive ablation studies confirm that each component contributes significantly to performance gains, and future work will explore more efficient structure transfer mechanisms.