Fast Detection of Duplicate Bug Reports using LDA-based Topic Modeling and Classification

A bug tracking system continuously monitors the status of a software environment, like an Operating System (OS) or a user application. Whenever it detects an anomaly situation, it generates a bug report and sends it to the software developer or maintenance center. However, the newly reported bug can be an already existing issue that was reported earlier and may have a solution in the master report repository. This condition brings an avalanche of duplicate bug reports, posing a big challenge to the software development life cycle. Thus, early detection of duplicate bug reports has become an extremely important task in the software industry. To address this issue, this work proposes a double-tier approach using clustering and classification, whereby it exploits Latent Dirichlet Allocation (LDA) for topic-based clustering, multimodal text representation using Word2Vec (W2V), FastText (FT) and Global Vectors for Word Representation (GloVe), and a unified text similarity measure using Cosine and Euclidean metrics. The proposed model is tested on the Eclipse dataset consisting over 80,000 bug reports, which is the amalgamation of both master and duplicate reports. This work considers only the description of the reports for detecting duplicates. The experimental results show that the proposed two-tier model achieves a recall rate of 67% for Top-N recommendations with 3 times faster computation than the conventional one-on-one classification model.