Bucher Sahyouni

Preprint - 2026

MuSICRec

A multi-view graph recommender integrating collaborative, sequential, and multimodal information through contrastive learning.

Recommender systemsContrastive learningMultimodal learningGraph neural networks
Download paper PDF View on arXiv

Overview

MuSICRec is a multi-view graph recommender that integrates collaborative, sequential, and multimodal information within a unified contrastive framework. It is designed around the observation that useful alternative views can be formed from interaction histories themselves rather than through artificial graph perturbations.

Problem

Cold-start and sparsity are often addressed with sequential, multimodal, or contrastive signals. Combining these sources can also introduce noise, particularly when graph augmentations remove useful behaviour semantics or when textual and visual features are not well aligned with collaborative preference.

Method

MuSICRec introduces a sequence-item view in which each user history is attention-pooled into a sequence node. Propagation over the resulting sequence-item graph injects sequential context and creates an organic contrastive view without hand-crafted graph perturbations. A frozen multimodal item-item graph contributes visual and textual information, while an ID-guided gate controls how much of each modality enters the item representation before propagation.

Key contributions

  • Sequences represented as nodes in a sequence-item graph, adding contextual signals to collaborative filtering.
  • Entity-level contrast between each user and their own sequence as an alternative to artificial augmentation.
  • ID-guided cross-modal calibration to limit misalignment and modality noise.
  • A broad evaluation spanning collaborative, sequential, multimodal, and contrastive recommenders.

Results / findings

Under a strict leave-two-out protocol on the Amazon Baby, Sports, and Electronics datasets, MuSICRec consistently outperformed the evaluated methods across all four model families. The gains were most pronounced for users with short interaction histories, supporting the intended role of sequence-item propagation in sparse and cold-start settings. Ablations verified the contribution of the model components, and the reported sensitivity analysis found stable behaviour across the tested graph weighting parameters.

Citation

Bucher Sahyouni, Matthew Vowels, Liqun Chen, and Simon Hadfield. “Sequences as Nodes for Contrastive Multimodal Graph Recommendation.” arXiv:2602.07208, 2026.