Labeled Subgraph Entropy Kernel
This work addresses efficiency and accuracy issues in graph similarity measurement for domains like bioinformatics and chemistry, representing an incremental improvement.
The paper tackles the computational complexity and lack of node-level information in existing entropy-based graph kernels by proposing a labeled subgraph entropy kernel, which outperforms state-of-the-art methods on real-world datasets.
In recent years, kernel methods are widespread in tasks of similarity measuring. Specifically, graph kernels are widely used in fields of bioinformatics, chemistry and financial data analysis. However, existing methods, especially entropy based graph kernels are subject to large computational complexity and the negligence of node-level information. In this paper, we propose a novel labeled subgraph entropy graph kernel, which performs well in structural similarity assessment. We design a dynamic programming subgraph enumeration algorithm, which effectively reduces the time complexity. Specially, we propose labeled subgraph, which enriches substructure topology with semantic information. Analogizing the cluster expansion process of gas cluster in statistical mechanics, we re-derive the partition function and calculate the global graph entropy to characterize the network. In order to test our method, we apply several real-world datasets and assess the effects in different tasks. To capture more experiment details, we quantitatively and qualitatively analyze the contribution of different topology structures. Experimental results successfully demonstrate the effectiveness of our method which outperforms several state-of-the-art methods.