Bucher Sahyouni

Preprint - 2026

MuSTRec

A unified recommender framework combining multimodal item structure with sequential preference modelling.

Recommender systemsSequential recommendationMultimodal learningGraph neural networks
Download paper PDF View on arXiv

Overview

MuSTRec (Multimodal and Sequential Transformer-based Recommendation) combines multimodal item structure with sequential preference modelling. It brings collaborative, content-based, and context-based filtering into a single framework, using item content and interaction structure to support learning under sparse feedback.

Problem

Sequential recommendation captures the temporal evolution of user behaviour, but the number of available interaction sequences is much smaller than the number of raw interactions. Multimodal information can provide richer item representations and direct relationships between similar items, yet it is often treated separately from sequential modelling.

Method

MuSTRec constructs a user-item bipartite graph and frozen item-item graphs from textual and visual features. Graph-derived item representations are assembled into user interaction sequences and passed to a transformer-like prediction head. A frequency-based self-attention module separates high- and low-frequency information to model short- and long-term preference. The graph and sequential components are trained together through shared representations.

Key contributions

  • A unified architecture connecting collaborative, multimodal, and sequential recommendation.
  • Textual and visual item-item graphs that enrich item representations beyond interaction data.
  • Frequency-based self-attention for modelling both short- and long-term user preferences.
  • A revised partitioning regime and analysis of the role of user embeddings in sequential recommendation.

Results / findings

Across the evaluated Amazon datasets, MuSTRec consistently outperformed the multimodal and sequential baselines, with improvements of up to 33.5%. Ablation studies confirmed the importance of the graph-derived embeddings. The paper also found that integrating user embeddings can produce substantial short-term metric gains on smaller datasets, while further experiments indicate that these gains arise from the interaction between graph propagation and collaborative training rather than user-specific embeddings alone.

Citation

Bucher Sahyouni, Matthew Vowels, Liqun Chen, and Simon Hadfield. “Multimodal Enhancement of Sequential Recommendation.” arXiv:2602.07207, 2026.