Entire Space Delayed Feedback with Cross‑Task Knowledge Distillation (ESDC) for Multi‑Task E‑commerce Recommendation

Background: In e‑commerce recommendation, the ultimate goal is transaction (GMV) which proceeds through exposure → click → purchase. Modeling conversion rate (CVR) faces four major issues: sample selection bias, data sparsity, delayed feedback, and knowledge inconsistency between CTR and CVR.

To tackle these, the authors propose the Entire Space Delayed Feedback with Cross‑Task Knowledge Distillation (ESDC) model. The architecture consists of a teacher network (based on PLE and enhanced with ESCM² and ESDF ideas) and a student network, linked by a distillation weight‑adjustment module.

Teacher network: Defines random variables X, Y, Z, C, D, E and partitions samples into click‑space and entire‑space sets. It contains four modules: CTR (standard MLP), CVR (adds post‑click behavior features), Delay (discretizes delayed conversion time), and IPS weighting to debias CTR influence on CVR. Loss functions combine CTR loss, CTCVR loss (with ESDF), and IPS loss.

Distillation weight adjustment: Uses uncertainty weighting based on KL divergence between twin CVR networks (dropout‑based) and a common‑frequency loss to down‑weight noisy soft‑labels. The final distillation loss integrates these weights.

Student network: Mirrors the teacher structure but removes unavailable online features and incorporates purchase‑sample weighting in the CTR module. Its loss combines the distilled CVR loss and standard CTR loss.

Results: The ESDC model, when deployed in Xiaomi’s homepage feed recommendation, achieved a 15 % increase in GMV compared with the baseline PLE + ESMM model.

Conclusion: Accurate CVR estimation is crucial for e‑commerce recommendation. The ESDC framework demonstrates how causal inference, cross‑task distillation, twin networks, and uncertainty‑aware weighting can jointly address the four challenges and substantially improve business metrics.