Unsupervised Learning Exploration for Duplicate Bug Report Detection

Bugs are inevitable as software complexity increases and bug reports are crucial for program maintenance. Due to the variety of language expressions, there will be duplicate bug reports where different reports correspond to the same bug. Existing methods for detecting duplicate bug reports are supervised, but manually labeling data is time-consuming and difficult because it requires expertise to identify the specific bug in the report. To reduce the dependence of current duplicate bug report detection methods on data labels, we expect to explore an unsupervised approach for detecting duplicate bug reports based on GCN. Firstly, since the bug report dataset has no prior graph information, we utilize the estimated distribution of connectivity between two disjoint sets in a bipartite graph to build a graph. Then, anchor points are introduced to address the problem that traditional graph autoencoders are not suitable for large-scale datasets. Finally, we conduct extensive experiments to evaluate the performance of our unsupervised approach. Our work shows that unsupervised approaches are feasible for duplicate bug report detection.