| My dissertation concentrates on two problems in mining discriminative subgraphs: how to efficiently identify subgraph patterns that discriminate two sets of graphs and how to improve discrimination power of subgraph patterns by allowing flexibility. To achieve high efficiency, I adapted evolutionary computation to subgraph mining and proposed to learn how to prune search space from search history. To allow flexibility, I proposed to loosely assemble small rigid graphs for structural flexibility and I proposed a label relaxation technique for label flexibility.;I evaluated how applications of discriminative subgraphs can benefit from more efficient and effective mining algorithms. Experimental results showed that the proposed algorithms outperform other algorithms in terms of speed. In addition, using discriminative subgraph patterns found by the proposed algorithms leads to competitive or higher classification accuracy than other methods. Allowing structural flexibility enables users to identify subgraph patterns with even higher discrimination power. |
