A simple yet effective baseline for non-attribute graph classification

Chen Cai, Yusu WangarXiv 2018 · arXiv

Abstract

Researchers from The Ohio State University have introduced Local Degree Profile (LDP), a linear-time graph representation that requires zero training. This simple baseline performs comparably to, and sometimes better than, state-of-the-art Graph Neural Networks (GNNs) and complex graph kernels on popular social network benchmarks. These findings suggest that current datasets may not be sufficiently challenging to justify the complexity of modern deep learning models for graph classification.

Introduction

The field of graph classification has recently shifted toward increasingly sophisticated methods, primarily divided into mathematical graph kernels and deep Graph Neural Networks (GNNs). However, as these models grow in complexity and computational cost, it becomes difficult to discern which components drive performance. Are improvements due to the expressive power of deep learning, or are benchmark datasets simple enough that basic structural statistics suffice? To address this, the authors developed the Local Degree Profile (LDP), the simplest meaningful baseline, to test the limits of current benchmarks.

The Local Degree Profile (LDP) Method

LDP relies exclusively on node degrees—the number of neighbors a node has—avoiding backpropagation or complex spectral theory. For every node $v$ in a graph, the method examines its "1-neighborhood" and calculates five statistics:

  1. Degree($v$): The node's own degree.
  2. Min: The minimum degree among its neighbors.
  3. Max: The maximum degree among its neighbors.
  4. Mean: The average degree of its neighbors.
  5. Standard Deviation: The variation in neighbor degrees.

These features are aggregated into histograms or empirical distribution functions to represent the entire graph. The resulting vectors are then fed into a standard Support Vector Machine (SVM).

Table 1: Complexity comparison of LDP against various graph kernels. LDP achieves linear time complexity O(E), making it significantly more scalable than traditional kernels.

Crucially, LDP extraction is a linear-time process, $O(E)$, where $E$ is the number of edges. This makes it as fast as reading the graph data itself, offering a massive efficiency advantage over kernels with $O(V^3)$ or higher complexity.

Experimental Results

The authors evaluated LDP on two categories of benchmarks: non-attributed social networks (e.g., IMDB, REDDIT) and attributed bioinformatics datasets (e.g., MUTAG, PROTEINS).

Performance on Social Networks

On social network datasets, LDP frequently matched or outperformed state-of-the-art GNNs like Graph Isomorphism Networks (GIN) and DiffPool. For instance, on the REDDIT-BINARY dataset, LDP achieved 92.1% accuracy, nearly identical to GIN’s 92.4%.

Table 2: Classification accuracy on social network datasets. LDP performs competitively with sophisticated GNN architectures like GIN and DiffPool.

Performance on Attributed Graphs

While LDP is designed for non-attributed graphs, it remains competitive on chemical and biological datasets when augmented with node labels. However, its performance is generally lower in these domains compared to specialized kernels, as local degree information alone is often insufficient for capturing complex molecular properties.

Discussion and Insights

The success of LDP offers several critical insights for the graph learning community:

  • GNNs as Feature Aggregators: LDP can be interpreted as a 2-layer GNN with fixed, non-trainable weights. Its high performance suggests that for many tasks, the structural information in the neighborhood is more critical than the learned weights.
  • Robustness: Unlike the Weisfeiler-Lehman (WL) kernel, which uses exact hashing and can be sensitive to minor graph perturbations, LDP’s use of continuous statistics makes it more robust.
  • Benchmark Limitations: The fact that a simple histogram of degrees can rival deep GNNs suggests that current social network benchmarks may be "too easy." They may primarily be distinguishable by local degree distributions rather than global structural patterns.

Conclusion

LDP serves as a powerful reminder of the importance of strong baselines. While it struggles with global structures—such as long-range paths—its efficiency and performance on standard benchmarks are remarkable. The authors call for the development of more diverse and complex "ImageNet-scale" graph datasets. Until such benchmarks are available, any new GNN architecture should be measured against LDP to ensure its complexity provides genuine value.

Summary updated

Abstract

Graphs are complex objects that do not lend themselves easily to typical learning tasks. Recently, a range of approaches based on graph kernels or graph neural networks have been developed for graph classification and for representation learning on graphs in general. As the developed methodologies become more sophisticated, it is important to understand which components of the increasingly complex methods are necessary or most effective. As a first step, we develop a simple yet meaningful graph representation, and explore its effectiveness in graph classification. We test our baseline representation for the graph classification task on a range of graph datasets. Interestingly, this simple representation achieves similar performance as the state-of-the-art graph kernels and graph neural networks for non-attributed graph classification. Its performance on classifying attributed graphs is slightly weaker as it does not incorporate attributes. However, given its simplicity and efficiency, we believe that it still serves as an effective baseline for attributed graph classification. Our graph representation is efficient (linear-time) to compute. We also provide a simple connection with the graph neural networks. Note that these observations are only for the task of graph classification while existing methods are often designed for a broader scope including node embedding and link prediction. The results are also likely biased due to the limited amount of benchmark datasets available. Nevertheless, the good performance of our simple baseline calls for the development of new, more comprehensive benchmark datasets so as to better evaluate and analyze different graph learning methods. Furthermore, given the computational efficiency of our graph summary, we believe that it is a good candidate as a baseline method for future graph classification (or even other graph learning) studies.
BibTeX
@misc{cai2018simple,
  author = {Chen Cai and Yusu Wang},
  title = {{A simple yet effective baseline for non-attribute graph classification}},
  year = {2018},
  eprint = {1811.03508},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1811.03508},
}
Figure