Predicting defect resolution time using cosine similarity

Defect resolution on time is one of the overriding project goals which cannot be neglected. Often projects suffer from missed deadlines due to open critical defects. This negatively impacts successful delivery of a product, resulting in loss of revenue and customer dissatisfaction. Predicting defect resolution time, though a daunting task, can alleviate this risk of missing targeted milestones. In this paper, the authors propose three supervised learning approaches leveraging cosine similarity measure, progressively improving the prediction for days to resolve (DTR) a defect. The prediction model uses historical defect data to estimate DTR for new similar defects. The first prediction approach leverages Naive Bayes Classifier (NBC) to assess project risks by answering: Is quicker defect resolution feasible? The outcome of this analysis gives preliminary information on the resolution duration. To gain deeper insights on DTR, second approach utilizes similarity score between two defect summaries to predict DTR. To improve the prediction accuracy further, a third approach is shown, where predictions are based on statistical analysis on DTR of defects having same similarity scores. This approach yields lower error rates in predicting DTR for P2-High and P3-Medium defects, as compared to the second approach. Both the approaches however outperforms the simple approach, not involving supervised learning. These approaches can be applied over both open and closed source projects to reduce defect DTR.