DaTaobao Tech
May 31, 2022 · Artificial Intelligence
Decoupling Popularity Bias in Dual‑Tower Retrieval Models
The paper proposes CDAN, a dual‑tower retrieval model that separates item attribute and popularity representations via a Feature Decoupling Module with orthogonal embeddings, aligns head‑tail attribute distributions using MMD and contrastive learning, and jointly trains biased and unbiased towers, achieving higher tail recall, lower exposure concentration, and measurable online click‑through improvements.
Recommendation systemscontrastive learningdomain adaptation
0 likes · 13 min read