Generalizing Weisfeiler-Lehman Kernels to Subgraphs
Dongkwan Kim, Alice Oh
TL;DR
This paper introduces WLKS, a generalized Weisfeiler-Lehman kernel for subgraphs that enhances structural representation by applying WL on k-hop neighborhoods, outperforming existing methods in accuracy and efficiency.
Contribution
The paper proposes WLKS, a novel kernel that extends WL to subgraphs using k-hop neighborhoods, improving expressiveness and efficiency over current GNN approaches.
Findings
WLKS outperforms leading methods on five datasets.
Reduces training time significantly, from 0.01x to 0.25x.
Effectively captures complex subgraph structures.
Abstract
Subgraph representation learning has been effective in solving various real-world problems. However, current graph neural networks (GNNs) produce suboptimal results for subgraph-level tasks due to their inability to capture complex interactions within and between subgraphs. To provide a more expressive and efficient alternative, we propose WLKS, a Weisfeiler-Lehman (WL) kernel generalized for subgraphs by applying the WL algorithm on induced -hop neighborhoods. We combine kernels across different -hop levels to capture richer structural information that is not fully encoded in existing models. Our approach can balance expressiveness and efficiency by eliminating the need for neighborhood sampling. In experiments on eight real-world and synthetic benchmarks, WLKS significantly outperforms leading approaches on five datasets while reducing training time, ranging from 0.01x to 0.25x…
Peer Reviews
Decision·ICLR 2025 Poster
- The paper is well written. - The method is simple: it is easy to understand while offering a compelling trade-off between complexity and performance. - Different aspects of the method are empirically analysed, while a couple of theoretical insights are offered. - The combination of the approach with GNN models feels slightly arbitrary, but nonetheless it can already provide initial insights on the relative importance of feature information, and shows how the approach is compatible with other m
- The manuscript would benefit from a preliminary section that would more formally describe the problem setting of subgraph representation learning, and the shortcomings of state-of-the-art approaches which the authors seek to overcome. - The paper lacks a more detailed, *formal* comparison with other state-of-the-art approaches. From a methodological standpoint, beyond experimental results, how is the method different than other approaches such as SubGNN? Is it capturing some information that t
1. WLKS improves subgraph-level task performance by encoding both internal subgraph structures and their k-hop neighborhoods, offering a more expressive alternative to traditional GNNs. 2. WLKS achieves superior performance with reduced training times, outperforming several state-of-the-art models by less training time, without requiring GPUs, pre-training, complex hyperparameter tuning, or extensive pre-computation. 3. WLKS can be integrated into GNN frameworks (like S2N) as an adjacency matr
1. WLKS depends on the node coloring algorithm, making it less suitable for datasets where nodes or edges have continuous feature values, which are challenging to map to discrete color values. It would be beneficial for the authors to discuss potential adaptations of WLKS for handling continuous node and edge features, or to clarify if existing techniques could address this limitation. 2. The selection of k in WLKS is somewhat limited, focusing solely on the original subgraph and the entire glo
W1. The research significance of subgraph learning in graph representation learning is undoubtedly valuable. W2. The research in this manuscript is interesting, and the proposed framework of both efficiency and effectiveness is exciting.
W1. The analysis of existing challenges is unconvincing, and this manuscript lacks an analysis of existing works. The author identifies the main challenge as capturing arbitrary interactions between and within subgraph structures. However, the local-global interactive learning strategy has been well studied by GNN-AK[1]. W2. This manuscript lacks sufficient interpretability analysis of the proposed method. For instance, the right panel of Figure 1 only shows the final coloring result, leaving
Code & Models
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
Topicsadvanced mathematical theories