Abstract:
Assignment of new issues to developers is an important part of software maintenance activities. In this research, we build an issue recommendation model that recommends new issues to developers and highlights the defect prone software modules to the developer who owns an issue. Existing solutions address the automated issue assignment problem through text mining. We propose a recommender model that uses the collaboration network of developers on software modules and the structured issue data as its input. We perform an exploratory analysis using the issue data of two large software systems and observe the trends in the issue ownership, issue defect relations and issue timelines. Our base model estimates the developer centrality for each issue category and recommends issues to developers based on their centrality ranking. We test the performance of our recommender using the maintenance data of a globally developed large enterprise software using recommendation accuracy and workload imbalance as the performance metrics. We extend the recommender to address, (i) The problem of developer workload imbalances, (ii) The problem of assigning issues to a new group of developers by using stochastic Kronecker networks to model the future structure of the collaboration network. We change a learning based defect predictor's output based on recent history to update the predicted defect-prone software modules on a real-time basis. Our empirical analysis shows that: (i) The performance of our recommender model approximates historical trends of issue allocation, (ii) Heuristics applied to the model output reduces the issue ownership inequalities and model approximation performance, (iii) Kronecker networks can be used to estimate the future collaboration on the issues and the model can be extended by them to propose issues to new developer groups, (iv) Real time defect prediction model can signi cantly improve probability of detection over time while not changing false alarm rates.
