- Rating
- 1
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes ReCoG, a framework for multimodal graph learning that jointly refines graph topology and learns multimodal node representations through end-to-end reciprocal interaction. It consists of a multimodal graph refiner that uses cross-modal semantic information to correct existing edges and identify potential missing connections, and a coupled cross-modal message passing mechanism that enables intra- and inter-modal information propagation over the refined graph. The paper also provides a spectral analysis of the coupled propagation operator and derives a contraction bound under specific conditions. Experiments on nine benchmarks for node classification and link prediction report improvements over a range of existing baselines.
Critical strengths
-
A new framework for multimodal graph learning. ReCoG jointly addresses graph topology refinement and multimodal representation learning, combining a multimodal graph refiner with coupled cross-modal message passing. This is a reasonable approach to handling noisy or incomplete graph structures while leveraging complementary modalities.
-
Broad empirical evaluation across tasks and datasets. The paper evaluates ReCoG on nine benchmarks covering node classification and link prediction, with multiple feature encoder configurations. The reported improvements across a range of baselines suggest that the proposed framework has practical potential beyond a single dataset or task.
Critical weaknesses
-
The novelty of the proposed framework appears limited. Components such as similarity-guided candidate selection, residual graph corrections, and cross-modal message passing closely parallel existing approaches in graph structure learning and multimodal fusion, making the overall methodological contribution somewhat incremental.
-
The multimodal fusion method introduced in Section 3.3 is a node-adaptive scalar gating mechanism. Specifically, for each node , the textual feature and visual feature are concatenated and passed through an MLP followed by a Sigmoid function to derive a scalar gate . This gate then modulates the weighted sum of both modalities before applying Layer Normalization. This formulation is conceptually identical to the Gated Multimodal Unit (GMU) proposed by Arevalo et al. (Gated Multimodal Units for Information Fusion, ICLR 2017 Workshop). The authors didn't cite this baseline work or discuss how their proposed fusion mechanism differs from or improves upon such standard gated fusion units.
Questions for authors
-
Could the authors explain why the multimodal graph refiner improves node classification and link prediction?
-
What enables the proposed multimodal graph refiner to achieve better results compared to existing refinement/GSL approaches you mentioned in the Related Work, such as IDGL (Chen et al., 2020b), SLAPS (Fatemi et al., 2021), ProGNN (Jin et al., 2020), and NodeFormer (Wu et al., 2022a)? I also suggest including a discussion and comparison with recent relevant literature from 2025–2026 to better position the paper within the current state of the art.