Relative Code Comprehensibility Prediction

Automatically predicting how difficult it is for humans to understand a code snippet can assist developers in tasks like deciding when and where to refactor. Despite many proposed code comprehensibility metrics, studies have shown they often correlate poorly with actual measurements of human comprehensibility. This has motivated the use of machine learning models to predict human comprehensibility directly from code, but these models have also shown limited accuracy. We argue that model inaccuracy stems from inherent noise in human comprehensibility data, which confuses models trained to predict it directly. To address this, we propose training models to predict the relative comprehensibility of two code snippets - that is, predicting which snippet a human would find easier to understand without predicting each snippet's comprehensibility in isolation. This mitigates noise in predicting 'absolute' comprehensibility measurements, but is still useful for downstream software-engineering tasks like assessing whether refactoring improves or hinders comprehensibility. We conducted a study to assess and compare the effectiveness of absolute and relative code comprehensibility prediction via machine learning. We used a dataset of 150 Java code snippets and 12.5k human comprehensibility measurements from prior user studies, comparing the models' performance with naive baselines (eg 'always predict the majority class'). Our findings indicate that absolute comprehensibility models improve over the baselines by at most 33.4% and frequently underperform. In contrast, relative comprehensibility models are substantially better, with average improvements of 137.8% and 74.7% for snippet-wise and developer-wise prediction, respectively. These results suggest that relative comprehensibility models learn more effectively from the data, supporting their practical applicability for downstream SE tasks.