# DSGRec: dual-path selection graph for multimodal recommendation

## TL;DR

This paper introduces DSGRec, a new recommendation system that improves accuracy by combining user behavior and multi-modal data through a dual-path graph architecture.

## Contribution

The novel dual-path selection architecture enhances collaboration between user behavior and multi-modal information in recommendation systems.

## Key findings

- DSGRec outperforms state-of-the-art methods on three benchmark datasets.
- The dual-path design improves modeling of user behavior and multi-modal signals.
- Contrastive learning tasks help align auxiliary signals with user-item interactions.

## Abstract

With the advancement of digital streaming technology, multi-modal recommendation systems have gained significant attention. Current graph-based multi-modal recommendation approaches typically model user interests using either user interaction signals or multi-modal item information derived from heterogeneous graphs. Although methods based on graph convolutional networks (GCNs) have achieved notable success, they still face two key limitations: (1) the narrow interpretation of interaction information, leading to incomplete modeling of user behavior, and (2) a lack of fine-grained collaboration between user behavior and multi-modal information. To address these issues, we propose a novel method by decomposing interaction information into two distinct signal pathways, referred to as a dual-path selection architecture, named Dual-path Selective Graph Recommender (DSGRec). DSGRec is designed to deliver more accurate and personalized recommendations by facilitating the positive collaboration of interactive data and multi-modal information. To further enhance the represetation of these signals, we introduce two key components: (1) behavior-aware multimodal signal augmentation, which extract rich multimodal semantic information; and (b) hypergraph-guided cooperative signal enhancement, which captures hybrid global information. Our model learns dual-path selection signals via a primary module and introduces two auxiliary modules to adjust these signals. We introduce independent contrastive learning tasks for the auxiliary signals, enabling DSGRec to explore the mechanisms behind feature embeddings from different perspectives. This approach ensures that each auxiliary module aligns with the user-item interaction view independently, calibrating its contribution based on historical interactions. Extensive experiments conducted on three benchmark datasets demonstrate the superiority of DSGRec over several state-of-the-art recommendation baselines, highlighting the effectiveness of our method.