TL;DR
GraSP is a simple yet powerful graph similarity prediction method that leverages enhanced features and advanced GNN architectures to outperform existing models in accuracy and efficiency.
Contribution
Introducing GraSP, a novel GSC approach that surpasses prior methods with improved features, gating, residuals, and multi-scale pooling, and theoretically exceeds the 1-WL test.
Findings
GraSP outperforms 10 competitors on benchmark datasets.
GraSP achieves higher effectiveness and efficiency.
Theoretical analysis shows GraSP surpasses 1-WL test.
Abstract
Graph similarity computation (GSC) is to calculate the similarity between one pair of graphs, which is a fundamental problem with fruitful applications in the graph community. In GSC, graph edit distance (GED) and maximum common subgraph (MCS) are two important similarity metrics, both of which are NP-hard to compute. Instead of calculating the exact values, recent solutions resort to leveraging graph neural networks (GNNs) to learn data-driven models for the estimation of GED and MCS. Most of them are built on components involving node-level interactions crossing graphs, which engender vast computation overhead but are of little avail in effectiveness. In the paper, we present GraSP, a simple yet effective GSC approach for GED and MCS prediction. GraSP achieves high result efficacy through several key instruments: enhanced node features via positional encoding and a GNN model augmented…
Peer Reviews
Decision·Submitted to ICLR 2024
1. The method is well-motivated with the observation that cross-graph node-level interaction is costly, and the proposed method is relatively simple with great inference time reduction. 2. The paper is well written and easy to follow.
1. It would be interesting to know what would be the alternative to the proposed combination of summation and attention pooling, e.g. what if a simple non-learning combination is used, i.e. $\alpha * z_{sum} + (1 - \alpha) * z_{att}$ where $\alpha$ is a hyperaprameter scalar. This will further highlight the importance of learnable combination of the two pooling methods. 2. The overall novelty is limited, given the adoption of random walk positional encoding and combination of existing graph poo
1. The paper focuses on predicting graph similarity/distance metrics which is a very important problem. The organization of the paper is good and easy to follow. 2. The authors used positional encoding to enhance node features which also aided in passing the 1-WL test. 3. The ablation study is good which covered most of the design choices the authors made.
1. The authors presented MSE on predicted similarity scores (obtained by exponentiating a normalized version of GED) instead of predicted GED. Given the task is to predict GED, results should be reported on GED itself and not its transformation to a similarity metric that distorts the true errors. While some previous works such as SIMGNN have also followed the same methodology of reporting results on exponentiated similarity instead of true GED, there is no justification for this transformation.
S1: The authors innovatively incorporate positional encoding within GNN framework, a commendable step that advances the GNN's ability to capture nuanced structural information. This ingenuity potentially sets a new precedent for subsequent research in graph similarity assessment.\ S2: The methodology introduced in this paper demonstrates notable efficiency. \ S3: The experimental results seem to be promising.
W1: The presentation of the content, particularly in Section 3, lacks clarity and cohesiveness, making it challenging for readers to follow and understand the proposed methodology. \ W2: The paper posits inefficiency in contemporary cross-graph interaction techniques as a primary catalyst for the development of GRASP. However, the narrative lacks a coherent demonstration of how GRASP mitigates these inefficiencies. The empirical section, intended to validate the method's enhanced efficiency, doe
Code & Models
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
