# RoleMAG: Learning Neighbor Roles in Multimodal Graphs

Source: [RoleMAG: Learning Neighbor Roles in Multimodal Graphs](<https://www.sota2.com/research/paper/69df1b6d7d0172140ec7b792>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-04-16T05:00:04.684Z

Authors: Yilong Zuo, Xunkai Li, Zhihan Zhang, Ronghua Li, Guoren Wang

Published: 2026-04-14T04:35:02.000Z

## Abstract

Multimodal attributed graphs (MAGs) combine multimodal node attributes with structured relations. However, existing methods usually perform shared message passing on a single graph and implicitly assume that the same neighbors are equally useful for all modalities. In practice, neighbors that benefit one modality may interfere with another, blurring modality-specific signals under shared propagation. To address this issue, we propose RoleMAG, a multimodal graph framework that learns how different neighbors should participate in propagation. Concretely, RoleMAG distinguishes whether a neighbor should provide shared, complementary, or heterophilous signals, and routes them through separate propagation channels. This enables cross-modal completion from complementary neighbors while keeping heterophilous ones out of shared smoothing. Extensive experiments on three graph-centric MAG benchmarks show that RoleMAG achieves the best results on RedditS and Bili\\\_Dance, while remaining competitive on Toys. Ablation, robustness, and efficiency analyses further support the effectiveness of the proposed role-aware propagation design. Our code is available at https://anonymous.4open.science/r/RoleMAG-7EE0/

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Node Classification on Bili\_Dance](<https://www.sota2.com/research/sota/node-classification-on-bili_dance.md>): Node Classification; Bili\_Dance; Accuracy: 44.82.

- [Node Classification on RedditS](<https://www.sota2.com/research/sota/node-classification-on-reddits.md>): Node Classification; RedditS; Accuracy: 92.57.

- [Link Prediction on Bili Dance](<https://www.sota2.com/research/sota/link-prediction-on-bili-dance.md>): Link Prediction; Bili Dance; MRR: 30.28.

- [Node Classification on Grocery](<https://www.sota2.com/research/sota/node-classification-on-grocery.md>): Node Classification; Grocery; Accuracy: 83.45.

- [Node Classification on Toys](<https://www.sota2.com/research/sota/node-classification-on-toys.md>): Node Classification; Toys; Accuracy: 78.83.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2604.12271>)

[PDF](<https://arxiv.org/pdf/2604.12271.pdf>)