SBFRec: Semantic-Behavioral Fusion with Trajectory Smoothing for Generative Sequential Recommendation

Abstract

Generative diffusion/flow-based models have recently shown strong potential for sequential recommendation, but their performance can be brittle when conditional signals from user histories are weak and semantic augmentation or routing is unreliable. In addition, common generative objectives may be misaligned with Top-K ranking behavior, which can exacerbate popularity bias under sparse and skewed interactions. We propose SBFRec, a Semantic-Behavioral Flow Matching framework for generative sequential recommendation. SBFRec adopts a stacked dual-branch encoder that combines a bidirectional RG-LRU behavioral branch for capturing sequential transition patterns with a sparse MoE semantic adapter for injecting pretrained priors, and fuses them through a shared residual stream with delta gating. We further introduce Popularity-Aware Adaptive Fusion (PAAF), a sequence-level routing mechanism that learns adaptive branch preference conditioned on the popularity profile, enabling the model to leverage the more reliable signal source for each sequence. The encoder conditions a flow-matching decoder for generative recommendation, and is trained with a multi-task objective augmented with dual-branch regularization to prevent branch collapse and encourage complementary signals from both branches. Experiments on four benchmark datasets show that SBFRec achieves an average 6.52% improvement over state-of-the-art methods on Top-K ranking metrics.