Bug-Fixing Time Prediction: Albert, Distilbert, and Google Bert Compared

Transformer-based models have increasingly been explored to automate bug report analysis. In this study, we examine the ability of three such models: ALBERT, DistilBERT, and Google BERT, to predict the time required to fix bugs, using a dataset collected from the LiveCode Bugzilla repository. We focus on four evaluation criteria: accuracy, F1-score, root mean squared error, and inference latency. During five experimental runs, ALBERT consistently outperformed the other two models, achieving accuracy 89% and an F1 score of 0.87, along with the lowest RMSE (2.10). Moreover, it reduced the inference time by approximately 25% compared to Google BERT. These results suggest that ALBERT strikes a promising compromise between prediction quality and computational efficiency, making it a practical candidate for integration into large-scale bug triaging systems