GraphTorque: Torque-Driven Rewiring Graph Neural Network
Sujia Huang, Lele Fu, Zhen Cui, Tong Zhang, Na Song, Bo Huang
TL;DR
GraphTorque introduces a novel torque-inspired hierarchical rewiring method for GNNs, dynamically adjusting graph structure to enhance message passing and improve learning in diverse graph types.
Contribution
It proposes a new torque-driven rewiring strategy based on classical mechanics principles, effectively improving GNN performance on heterophilous and homophilous graphs.
Findings
Outperforms existing rewiring methods on benchmark datasets.
Effective in both heterophilous and homophilous graph scenarios.
Enhances message passing by pruning high-torque edges and adding low-torque links.
Abstract
Graph Neural Networks (GNNs) have emerged as powerful tools for learning from graph-structured data, leveraging message passing to diffuse information and update node representations. However, most efforts have suggested that native interactions encoded in the graph may not be friendly for this process, motivating the development of graph rewiring methods. In this work, we propose a torque-driven hierarchical rewiring strategy, inspired by the notion of torque in classical mechanics, dynamically modulating message passing to improve representation learning in heterophilous and homophilous graphs. Specifically, we define the torque by treating the feature distance as a lever arm vector and the neighbor feature as a force vector weighted by the homophily disparity between nodes. We use the metric to hierarchically reconfigure receptive field of each layer by judiciously pruning…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
- The overall writing of the paper is relatively clear. However, it is worth noting that the citation format is used incorrectly throughout the entire paper, which significantly affects readability. - Determining the optimal graph connectivity based on the given node features is inherently a very challenging problem. - The authors provide experimental results on datasets with large-scale edges.
- Modeling node feature similarity and label consistency as a concept of “torque” lacks strong justification. In fact, the authors merely intend to express the combined influence of these two metrics on edge probabilities. There are, in practice, many ways to model such combined effects. The choice of using torque for this purpose is not well justified and is therefore unconvincing. - In line 274, the authors mention selecting the top‑t most similar nodes as candidate nodes. Theoretically, this
- This is the first work that uses a physics-inspired torque idea as a graph rewiring technique; this is novel and well motivated by the successful application of torque in other fields. - Experiments show strong performance across different evaluation setups. - Design choices are thoroughly ablated.
- The method’s main limitation lies in its relatively higher computational complexity, primarily due to the need to sort edges after torque computation in order to identify the torque gap for edge removal. Of course, this could not be a limiting factor in practice, but the authors do not provide a runtime comparison. I would highly recommend a runtime comparison be included in the Supplementary. - The authors have mainly performed experiments on node-level dataset. However, a lot of the works on
- The paper presents an integrated, train-time rewiring mechanism instead of doing static preprocessing. - The method performs rewiring separately at each layer, which is an important design choice albeit not novel. - It includes ablations that justify each component of the method.
1. The motivation for this rewiring method is lacking. In parts of the paper, "bad" edges are described as spurious or missing, but in the experiments the method is compared against approaches targeting over-squashing, which is a different problem. 2. The paper relies almost entirely on the outdated assumption that "heterophily = bad" and does not present any alternative conceptual motivation or new theoretical insight beyond that. 3. The claim that adversarial edges tend to connect low-similari
* The rewiring method combines (i) the distance between node representations and (ii) the neighbor’s feature strength, weighted by the local homophily gap a principled and interesting criterion. * Unlike curvature or spectral based rewiring, the scoring is directly dependent on the task and the data, leveraging node features together with homophily information. * The model’s computational complexity appears operational on large graphs, making it practical beyond small benchmarks.
- **Ambiguity in the introduction on the heterophily definition.** The introduction conflates label heterophily with feature dissimilarity. Phrases such as “addressing heterophily, where nodes with dissimilar labels or features tend to be connected” and “heterophilous graphs typically connect node pairs with low similarity” assume that heterophilous edges usually join feature dissimilar nodes, whereas one can have similar features but different labels. This ambiguity harms the motivation o
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAdvanced Graph Neural Networks · Functional Brain Connectivity Studies · Graph Theory and Algorithms