Learning Developer‐Code Contribution Relationship to Improve Code Readability Assessment

ABSTRACT Automated code readability assessment is essential for software maintenance and evolution. Current approaches mainly rely on structural, syntactic, and semantic features of individual code files to perform assessment, yet they ignore repository‐level relational information. A critical missing dimension is the contribution relationships between developers and the code files they modify within a repository, which reflect individualized coding style consistency across files. To mitigate this research gap, we formally define and model the developer‐code contribution relationships (DCR) for code readability assessment. We first build a heterogeneous graph, where nodes represent developers and code files, and weighted edges quantify developers' contribution ratios. On top of this graph, we adopt a heterogeneous graph attention network (HAN) to learn relational embeddings, which capture cross‐file readability similarities caused by common developers. We then combine the learned relational embeddings with high‐quality features from state‐of‐the‐art pretrained code readability models to complete the readability assessment. Evaluated on a real GitHub repository dataset, our method delivers stable performance improvements over existing baselines. Extensive experiments also verify that the DCR feature maintains strong generalizability across various graph neural network architectures. This study demonstrates that exploiting repository‐level developer‐code relations is a promising direction to further boost the performance of automated code readability assessment.