Preference-oriented Diversity Model Based on Mutual Information for E-commerce Search Re-ranking (SIGIR 2024)

Re‑ranking in e‑commerce search aims to reorder items by considering both relevance and diversity, but existing methods often sacrifice one for the other.

This paper proposes PODM‑MI, a Preference‑oriented Diversity Model based on Mutual Information, which captures users’ evolving diversity preferences using a variational multi‑dimensional Gaussian representation (PON) and enhances preference‑item consistency via a mutual‑information‑driven module (SAM).

The model builds a utility matrix from the enhanced consistency to adaptively adjust rankings, balancing accuracy and diversity. Optimization combines a PRM classification loss with a mutual‑information loss, derived via variational inference.

Extensive online A/B tests on JD.com’s main search engine show significant gains in user conversion (UCVR) and result diversity, with visualizations (entropy, T‑SNE) confirming that the model aligns rankings with user intent across diverse scenarios.

Future work includes finer‑grained feature engineering, further improvements to intent modeling, and explicit influence of intent on ranking.

The team also announces recruitment opportunities in generative retrieval and ranking, inviting interested candidates to apply.