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/
Figures & tables
Figure 1. Illustration of the key motivation behind RoleMAG. Existing multimodal graph methods usually adopt a unified neighbor propagation pattern, implicitly treating all neighbors as if they contributed in a similar way. In practice, however, neighbors may play different roles in multimodal propagation: some provide shared support, some serve as complementary neighbors that enable cross-modal completion, while others correspond to heterophilous relations that should not be mixed into shared smoothing. This motivates a fine-grained, role-aware view of neighborhood propagation.
Figure 2. Empirical observations behind RoleMAG. (a) Neighbor utility is modality-dependent: text and image branches prefer different neighborhood organizations. (b) When text representations are degraded, complementary completion is most effective along the I→T direction. (c) When image representations are degraded, the preferred direction reverses to T→I . (d) As heterophily increases, directly mixing all neighbors into shared smoothing becomes less reliable, whereas a dedicated heterophily-aware propagator remains more stable.
Figure 3. Framework of RoleMAG. RoleMAG performs role-aware multimodal propagation in three stages. First, an edge role router infers edge roles from semantic cues and uses structural features to calibrate routing strength. Second, the routed neighbors are sent to three specialized experts, including a shared expert for consistent smoothing, a directional complementary expert for cross-modal completion, and a heterophily expert for discrepant neighborhood interactions. Finally, the expert outputs are integrated by residual gating fusion and used for downstream tasks such as node classification, link prediction, and retrieval.
Model
Toys (Node Classification)
RedditS (Node Classification)
Bili_Dance (Link Prediction)
ACC
F1
ACC
F1
MRR
Hits@3
GCN
77.24 ±0.31
73.18 ±0.42
91.48 ±0.24
84.68 ±0.38
14.25 ±0.41
15.23 ±0.57
GAT
77.29 ±0.34
72.92 ±0.49
91.45 ±0.27
85.65 ±0.41
24.19 ±0.88
25.42 ±1.06
MMGCN
77.60 ±0.29
75.03 ±0.45
90.44 ±0.36
82.85 ±0.56
8.64 ±0.33
8.55 ±0.39
MGAT
78.12 ±0.42
74.21 ±0.61
91.73 ±0.31
85.12 ±0.48
16.84 ±0.51
18.92 ±0.72
LGMRec
78.81 ±0.73
72.31 ±1.14
92.36 ±0.19
86.10 ±0.32
11.46 ±0.37
11.78 ±0.44
Table 1. Main results on graph-centric tasks. All metrics are reported as percentages. The best, second best, and third best results are highlighted in dark red, royal blue, and orange, respectively.
Variant
Toys ACC
RedditS ACC
Shared-only
75.23
89.73
w/o Role Routing
77.23
90.85
w/o Complementary Expert
77.85
91.44
w/o Complementary Direction
77.93
91.62
w/o Heterophily Expert
76.97
90.44
Full RoleMAG
78.83
92.57
Table 2. Ablation study of RoleMAG on node classification. All metrics are reported as percentages.
Figure 4. Robustness and efficiency analysis.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Symbol
Meaning
Shape / Range
Input graph and node representations
G=(V,E,{X(m)}m∈M)
Multimodal attributed graph
–
M={T,I}
Set of modality indices, representing text and image, respectively
∣M∣=2
xi(m),X(m)
Modal input features of node vi and the modal feature matrix of the entire graph
Rdm , RN×dm
A,D,A~
Adjacency matrix, degree matrix, and symmetric normalized adjacency matrix
RN×N
N(i)
First-order neighborhood of node vi
⊆V
Appendix
Table 3. Notation used in Appendix B. The notations above are used consistently throughout this appendix. For detailed definitions of each variable, please refer to sections B.1 , B.2 , B.3 , B.4 , B.5 and B.6 .
Dataset
Task
#Nodes
#Edges
#Classes
Image Dim.
Text Dim.
Toys
Node Classification
20,695
63,443
18
768
768
RedditS
Node Classification
15,894
283,080
20
768
768
Bili_Dance
Link Prediction
2,307
9,127
–
768
768
Appendix
Table 4. Datasets used in the main-paper graph-centric experiments. Bili_Dance is used for link prediction and therefore does not involve node classes.
Component
Symbol / Hyperparameter
Value Used in Current Experiments
Node representation dimension
d
256
Dropout
p
0.2
Router hidden dimension
drouter
64
Complementary TopK size
K
16
Number of query tokens
Q
4
Attention bias scale
λbias
1.0
Appendix
Table 5. Key hyperparameters of RoleMAG used in the main experiments.
Item
Specification
Operating system
Ubuntu 22.04.5 LTS
Python
3.10.18
PyTorch
2.8.0
CUDA runtime
12.8
Graph libraries
PyG 2.6.1, DGL 2.4.0
GPU
NVIDIA RTX PRO 6000 Blackwell Server Edition
Appendix
Table 6. Main software and hardware environment used in the experiments.
Dataset
RoleMAG
DGF
LGMRec
Graph4MM
Acc
F1
Acc
F1
Acc
F1
Acc
F1
Toys
78.83
75.49
77.68
71.46
78.81
72.31
78.91
75.62
RedditS
92.57
87.04
91.73
82.63
92.36
86.10
92.34
86.31
Grocery
83.45
73.56
83.02
72.56
84.12
76.83
83.57
75.92
Bili_Dance
44.82
39.14
41.73
36.21
43.90
37.88
45.21
39.87
Appendix
Table 7. Additional node classification results on four MAG datasets. All metrics are reported in %.
Sweep
Value
Acc
F1
TopK
8
93.49
87.10
TopK
16
92.92
86.25
TopK
32
93.14
86.68
Queries
2
93.81
87.27
Queries
4
92.64
85.75
Queries
8
93.14
86.60
Appendix
Table 8. Hyperparameter sensitivity of structural interaction parameters on RedditS node classification. All metrics are reported in %.
Sweep
Value
Acc
F1
λqca
0.0
92.74
85.79
λqca
0.2
92.70
86.20
λqca
0.5
93.21
86.07
λqca
1.0
92.67
86.70
λbal
0.0
92.61
85.56
λbal
0.05
93.11
86.69
Appendix
Table 9. Hyperparameter sensitivity of auxiliary objective weights on RedditS node classification. All metrics are reported in %.
Reproducibility Item
Specification
Reference
Execution entrypoint
The graph-centric experiments are launched through python src/main.py , where the three fields task , dataset , and model jointly determine a run configuration.
sections D.6 and E.2
Environment specification
The dependency set is frozen by the environment file at the repository root, and the main software and hardware environment is listed in the environment table.
sections D.6 and 6
Task scope
The main paper covers two node classification tasks (Toys and RedditS) and one link prediction task (Bili_Dance).
sections D.1 and E.1
Data split protocol
Node classification uses a 60/20/20 node split, while link prediction uses a 70/10/20 edge split and is evaluated with ranking metrics.
sections D.1 and D.3
Input construction
Node inputs are constructed from frozen multimodal embeddings together with task-specific graph preprocessing.
section D.2
Model instantiation
RoleMAG is integrated into the training and evaluation workflow through the unified benchmark interface and instantiated along the router-expert-fusion pipeline specified in this paper.
sections D.4 and E.4
Appendix
Table 10. RoleMAG supplementary reproducibility checklist. The listed items specify the execution interface, configuration conventions, and output fields underlying the graph-centric experiments in the main paper.
Multimodal Attributed Graphs (MAGs) model real-world entities by coupling graph topology with heterogeneous attributes such as text and images. They support graph-centric tasks requiring structural and class-discriminative representations, and modality-centric tasks requiring fine-grained cross-modal correspondence. However, existing MAG methods often rely on fixed graph contexts or uniformly fused representations, causing task-agnostic propagation and over-compressed fusion that hinder diverse task requirements and modality-specific evidence preservation. To address this, we propose CoMAG, a unified MAG backbone that learns task-adaptive reliable contexts and modality-preserving alignment within them. CoMAG first conducts Reliable Context Learning by estimating edge reliability from multimodal semantic consistency, complementing raw topology with semantic neighbors, and selecting context components through a task-aware gate. It then performs Modality-preserving Hop-token Alignment by maintaining modality-specific multi-hop trajectories, matching modality-hop tokens across modalities, and decoupling shared and private representations. Thus, CoMAG produces graph and modality representations from one forward pass while retaining modality-specific cues. We further analyze stable propagation, over-smoothing mitigation, and modality-collapse control. Experiments on nine OpenMAG datasets compare CoMAG with feature-only, graph-only, multimodal, and unified MAG baselines across graph-level prediction, modality matching, and graph-conditioned generation. Results show that CoMAG achieves the best reported performance, demonstrating that task-adaptive reliable contexts and modality-preserving alignment improve structural prediction, cross-modal matching, and graph-conditioned generation while retaining sparse edge-linear complexity.
Multimodal-attributed graphs (MAGs) are a fundamental data structure for multimodal graph learning (MGL), enabling both graph-centric and modality-centric tasks. However, our empirical analysis reveals inherent topology quality limitations in real-world MAGs, including noisy interactions, missing connections, and task-agnostic relational structures. A single graph derived from generic relationships is therefore unlikely to be universally optimal for diverse downstream tasks. To address this challenge, we propose Task-aware Modality and Topology co-Evolution (TMTE), a novel MGL framework that jointly and iteratively optimizes graph topology and multimodal representations toward the target task. TMTE is motivated by the bidirectional coupling between modality and topology: multimodal attributes induce relational structures, while graph topology shapes modality representations. Concretely, TMTE casts topology evolution as multi-perspective metric learning over modality embeddings with an anchor-based approximation, and formulates modality evolution as smoothness-regularized fusion with cross-modal alignment, yielding a closed-loop task-aware co-evolution process. Extensive experiments on 9 MAG datasets and 1 non-graph multimodal dataset across 6 graph-centric and modality-centric tasks show that TMTE consistently achieves state-of-the-art performance. Our code is available at https://anonymous.4open.science/r/TMTE-1873.
Yinlin Zhu, Xunkai Li, Di Wu +3
Sun Yat-sen University, Guangzhou, China · Beijing Institute of Technology, Beijing, China
Multimodal Graph Neural Networks (MGNNs) have shown strong potential for learning from multimodal attributed graphs, yet most existing approaches rely on tightly coupled architectures that suffer from prohibitive computational overhead. In this paper, we present a systematic empirical analysis showing that decoupled MGNNs are substantially more efficient and scalable for large-scale graph learning. However, we identify a critical bottleneck in existing decoupled pipelines, namely modal conflict, which arises in both the propagation and aggregation stages. Specifically, independent multi-hop diffusion causes cross-modal semantic divergence during propagation, while naive fusion fails to align multi-hop feature trajectories during aggregation, jointly limiting effective representation learning. To address this challenge, we propose CAMPA, a Cross-modal Aligned Multimodal Propagation & Aggregation framework for decoupled multimodal graph learning. Concretely, CAMPA introduces a two-stage alignment mechanism: (1) cross-modal aligned propagation, which injects cross-modal similarity priors into message passing to preserve semantic consistency without additional parameter overhead; (2) trajectory aligned aggregation, which leverages trajectory-level self-attention and cross-attention to capture and align long-range dependencies across modalities and hops. Extensive experiments on diverse benchmark datasets and tasks demonstrate that CAMPA consistently outperforms strong coupled and decoupled baselines while preserving the efficiency advantages of the decoupled paradigm.