Finding software fault relevant subgraphs a new graph mining approach for software debugging

In this paper, a new approach for analyzing program behavioral graphs to detect fault suspicious subgraphs is presented. The existing graph mining approaches for bug localization merely detect discriminative subgraphs between failing and passing runs, which are not applicable when the context of a failure is not appeared in a discriminative pattern. In our proposed method, the suspicious transitions are identified by contrasting nearest neighbor failing and passing dynamic behavioral graphs. The technique takes advantage of null hypothesis testing and a new formula for ranking edges is presented. To construct the most bug relevant subgraph, the high ranked edges are applied and presented to the debugger. The experimental results on Siemens test suite and Space program reveal effectiveness of the proposed method on weighted dynamic graphs for locating bugs in comparison with other methods.