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

Abstract: Re‑ranking rearranges items by considering inter‑item relationships to better satisfy user needs. Existing methods improve scoring accuracy at the expense of diversity, or boost diversity while hurting accuracy. PODM‑MI addresses this trade‑off by using a variational inference‑based multivariate Gaussian to capture uncertain user diversity preferences and maximizing mutual information between these preferences and candidate items, yielding an adaptive utility matrix that balances relevance and diversity, with significant improvements demonstrated on JD's main search.

1. Background and Current Situation Users exhibit different decision stages (browsing, buying) with varying diversity needs, yet current models do not directly capture the relationship between decision stages and diversity. Re‑ranking must adapt to user demand: when diversity is needed, results should contain varied items; when accuracy is needed, results should focus on the most relevant category.

Challenges include accurately modeling evolving user intent and strengthening the match between results and this evolving intent.

To tackle these, the authors introduce PODM‑MI.

2. PODM‑MI

The model takes a ranked list and user behavior data (clicks, cart adds) as input. First, a Preference‑Oriented Network (PON) captures user diversity preferences and candidate item diversity representations. Then, a Similarity Alignment Module (SAM) enhances consistency between user preferences and item diversity. From this enhanced consistency, a utility matrix is derived to dynamically adjust rankings.

2.1 PON – User Preference Modeling Historical queries and associated items provide valuable signals of user intent. The method incorporates click streams, cart actions, and query trajectories. Unlike static embeddings, a multidimensional Gaussian (mean vector and diagonal covariance) models the evolution of user preferences, allowing uncertainty and measuring convergence/divergence via variance.

2.2 SAM – Optimizing Rankings with Mutual Information After modeling preferences and item diversity, mutual information quantifies the correlation between them. Maximizing this mutual information aligns the distribution of candidate items with user intent. Because direct estimation is intractable, a variational posterior estimator provides a tractable lower bound. A learnable utility matrix, obtained by dot‑product of learnable weights and alignment features, is multiplied with backbone scores to produce final rankings.

2.3 Optimization Objective and Final Loss

The overall loss combines a classification loss (prm loss) and a mutual‑information loss, guiding the model to balance relevance and diversity.

2.4 Experimental Results and Visualization Online A/B tests on JD’s e‑commerce search show that PODM‑MI increases user purchase likelihood and item diversity. Visualizations using t‑SNE on reduced query embeddings reveal distinct layers of divergent and convergent query flows, correlating with entropy of ranking results. Case studies demonstrate that for highly diverse query histories, PODM‑MI yields more varied results, while for convergent queries it produces more focused rankings.

3. Future Directions

Incorporate finer‑grained features to better model browsing‑buying intent.

Further optimize user intent modeling updates.

Make user intent modeling explicitly influence ranking decisions.

For questions or collaboration, contact {wanghuimu1, limingming65}@jd.com.

The JD Search Algorithm team is hiring for both full‑time and internship positions; interested candidates are encouraged to apply.

Recent related works include:

Breaking the Hourglass Phenomenon of Residual Quantization: Enhancing the Upper Bound of Generative Retrieval (arXiv:2407.21488)

Generative Retrieval with Preference Optimization for E‑commerce Search (arXiv:2407.19829)

A Preference‑oriented Diversity Model Based on Mutual‑information in Re‑ranking for E‑commerce Search (SIGIR 2024 Accepted)

MODRL‑TA: A Multi‑Objective Deep Reinforcement Learning Framework for Traffic Allocation in E‑commerce Search (CIKM 2024 Accepted)

Optimizing E‑commerce Search: Toward a Generalizable and Rank‑Consistent Pre‑Ranking Model (SIGIR 2024 Accepted)

Guest speakers: Dr. Wang Huimu (Institute of Automation, Chinese Academy of Sciences) and Dr. Li Mingming (Institute of Information Engineering, Chinese Academy of Sciences), both working on large models and retrieval at JD.