Aligning Multimodal Sequential Recommendations via Robust Direct Preference Optimization with Sparse MoE

TL;DR

This paper improves multimodal sequential recommendation by modifying Direct Preference Optimization with stochastic negative sampling, enhancing ranking performance while maintaining efficiency.

Contribution

It introduces a stochastic negative sampling strategy for DPO, combined with a sparse MoE encoder, leading to better ranking results in recommendation systems.

Findings

Replacing hard negatives with stochastic sampling improves ranking.

The method achieves up to 5.25% NDCG@5 on Amazon benchmarks.

The approach maintains inference efficiency while enhancing performance.

Abstract

Preference-based alignment objectives have been widely adopted, from RLHF-style pairwise learning in large language models to emerging applications in recommender systems. Yet, existing work rarely examines how Direct Preference Optimization (DPO) behaves under implicit feedback, where unobserved items are not reliable negatives. We conduct systematic experiments on multimodal sequential recommendation to compare common negative-selection strategies and their interaction with DPO training. Our central finding is that a simple modification, replacing deterministic hard negatives with stochastic sampling from a dynamic top-K candidate pool, consistently improves ranking performance. We attribute its effectiveness to two factors: (1) reducing erroneous suppressive gradients caused by false negatives, and (2) retaining informative hard signals while smoothing optimization via controlled stochasticity. With an optional sparse Mixture-of-Experts encoder for efficient capacity scaling, RoDPO achieves up to 5.25% NDCG@5 on three Amazon benchmarks, with nearly unchanged inference cost.