SPBC: A self-paced learning model for bug classification from historical repositories of open-source software

Abstract

Abstract One of the areas most in need of improvement in the field of automated bug fixing, localization and triaging systems is that of an effective categorization, as this would bugs to reduce the time, cost and effort required to locate, assign and fix the bug. The existing approaches depend upon the textual similarity of the bug description and category in a given reported bug; accordingly, the challenges of unstructured bugs, technical terms, versatile ways of reporting the same bug, the diverse nature and sizes of datasets etc. are often overlooked. Consequently, this limits the classifier performance to a specific type of dataset, resulting in classification inefficiency. To this end, we propose a novel Self-Paced Bug Classifier (SPBC) that is capable of locating the target categories from the bug description of the historical data, maintained by multiple open-source software packages (Bugzilla, Mentis, Redmine). The proposed model introduces a self-paced back-traceable algorithm, controlled by a self-paced regularizer, which classifies textually independent bug descriptions with weighted data-independent tokens (the easy samples). Later on, the regularizer sets comparatively hard samples for textually dependent classification by capturing intra-class and inter-class discrimination features from bug descriptions, based on the weighted similarities of words; this is done with the help of a Key Feature Identification Matrix (KFIM), a Non-Independent and Identically Distributed (NIID) matrix. Easy-to-hard self-pace learning, integrated with textually dependent and independent classification, makes SPBC capable of simultaneously enhancing the effectiveness and robustness of intelligent systems through a substantial increase in precision (5 to 15 % on average). The main advantage of SPBC is that it targets the spatial relationship between the data and the system, which makes it an apt learner of data and allows it to maintains sample insertion into the classifier at a controlled pace. Additionally, it maintains stability, which is not affected by the dataset’s dimensionality and traits. As is evidenced by the experimental results on four different datasets from open-source projects, our model outperforms the baseline and state-of-the-art methods through a single-stroke solution with improved accuracy and stable performance (average 95 % precision and 4% decrease in kappa); hence, it is significant for improving intelligent bug fixing and triaging systems.