Towards Combining Structured Pattern Mining and Graph Kernels

This paper presents a novel approach to feature construction for structured data in order to enhance graph prediction classification performance. To this end we combine graph mining techniques with graph kernel based classifiers. The main idea is to employ efficient mining techniques to extract a set of patterns correlated with the target concept and use these, or a selected subset of these, to annotate the original graph structures. A decomposition kernel is then defined on the enriched structured data instances. Experimental results on carcinogenic and toxicological activity prediction tasks for small molecules show that the proposed technique significantly increases classification performance.