Bidirectional encoder representations from transformers (BERT) driven approach for identifying feasible software enhancements

Software application users often submit numerous enhancement reports (ERs) requesting new features, but only a small fraction are deemed feasible and approved. Recent methods have attempted to automate the identification of feasible ERs, primarily using textual descriptions and traditional machine learning (ML) techniques. This article proposes a deep learning (DL) approach that significantly improves performance by integrating both textual (summary, sentiment) and non-textual metadata (reporter and module statistics) into a unified representation. Textual features are encoded using Bidirectional Encoder Representations from Transformers (BERT), while sentiment is computed using Senti4SD, a tool designed for sentiment analysis in software engineering. Non-textual features are derived from reporter and module histories to provide behavioral context. These features are concatenated and fed into a DL-based binary classifier. Experiments on a publicly available dataset show that the proposed approach substantially outperforms previous methods, improving accuracy from 82.38% to 94.02% and F1-score from 85.03% to 94.26%. The results highlight the effectiveness of combining semantic, affective, and behavioral features in predicting feasible ERs.