Abstract
Previous studies that used data from Stack Overflow to develop predictive models often employed limited benchmarks of 3–5 models or adopted arbitrary selection methods. Despite being insightful, their limited scope suggests the need to benchmark more models to avoid overlooking untested algorithms. Our study evaluates 21 algorithms across three tasks: predicting the number of question a user is likely to answer, their code quality violations, and their dropout status. We employed normalisation, standardisation, as well as logarithmic and power transformations paired with Bayesian hyperparameter optimisation and genetic algorithms. CodeBERT, a pre-trained language model for both natural and programming languages, was fine-tuned to classify user dropout given their posts (questions and answers) and code snippets. We found Bagging ensemble models combined with standardisation achieved the highest R2 value (0.821) in predicting users’ answers. The Stochastic Gradient Descent regressor, followed by Bagging and Epsilon Support Vector Machine models, consistently demonstrated superior performance to other benchmarked algorithms in predicting users’ code quality across multiple quality dimensions and languages. Extreme Gradient Boosting paired with log-transformation exhibited the highest F1-score (0.825) in predicting users’ dropout. CodeBERT was able to classify users’ dropout with a final F1-score of 0.809, validating the performance of Extreme Gradient Boosting that was solely based on numerical data. Overall, our benchmarking of 21 algorithms provides multiple insights. Researchers can leverage findings regarding the most suitable models for specific target variables, and practitioners can utilise the identified optimal hyperparameters to reduce the initial search space during their own hyperparameter tuning processes.