Retrieval-augmented diffusion with structural uncertainty for sequential recommendation

Dongjin Yu

doi:10.1016/j.eswa.2026.133588

Retrieval-augmented diffusion with structural uncertainty for sequential recommendation

2026-07-01

0

PRE

AI

DOI：10.1016/j.eswa.2026.133588

Original

Original request for help

Save

Abstract

En 中文

Sequential recommendation predicts users' next interactions based on their behavioral history but often suffers from data sparsity. Recently, contrastive learning has emerged as a promising solution by generating augmented views. However, existing methods typically adopt a uniform strategy across all user sequences. This one-size-fits-all strategy overlooks the structural uncertainty of user interests, where some sequences are robust under augmentation while others are vulnerable to semantic drift. To bridge this gap, we propose Retrieval-Augmented Diffusion for Sequential Recommendation (RADRec). Specifically, we first introduce Structural Entropy and Coupling Estimation to quantify the diversity and connectivity of user interests, respectively. Guided by these dual metrics, RADRec identifies sequences with competing and weakly connected interests as structurally vulnerable, for which we adopt self-alignment to mitigate semantic inconsistency. Conversely, other sequences are treated as structurally robust patterns, where we utilize a retrieval-guided diffusion model to synthesize complementary views for contrastive learning. Extensive experiments on four public datasets demonstrate that RADRec outperforms state-of-the-art baselines, learning more robust representations even under noisy and sparse data conditions.

Keywords:

Sequential recommendation

Contrastive learning

Semantic drift

Diffusion models

Retrieval-augmented diffusion with structural uncertainty for sequential recommendation

Abstract

Journal

Organization

Cited Papers