Multi-Granularity Community Smell Prediction with Developer-Oriented Features and Process Metrics
community smell; developer sentiment; socio-technical analysis; empirical software engineering
Open-source software development is a social activity organized by communities. Recent studies show that OSS quality and productivity are shaped by both technical and social factors, yet social debt — the extra cost caused by low-cohesion communication and collaboration — has received far less attention than its technical counterpart. Community smells are unhealthy organizational structures that operationalize social debt. Empirical evidence shows they significantly hamper refactoring and accelerate software aging. Prior work targets community managers and refactors communities top-down; however, OSS communities are flatter and more heterogeneous than traditional hierarchies, so a purely top-down lens is insufficient.
Community smells originate in development activities and are directly driven by developers. We therefore consider individual developers' experience, sentiment, and process characteristics, aiming to prevent community smells at the individual level. We examine whether these features suffice for prediction, and interpret their influence from both individual and community perspectives.
We build machine-learning models over experience, sentiment, and process features to predict whether a developer is affected by three prevalent community smells (lone wolf, organizational silo, and bottleneck), and whether an affected developer subsequently leaves the community. At the individual granularity, we apply SHAP to well-performing models to quantify feature importance and effect direction. At the community granularity, we forecast the number of affected developers to support maintainers' restructuring decisions.
Under within-project, time-sensitive, and cross-project validation, the individual-level models achieve F1 scores of 0.73–0.92; the community-level forecast attains an R² of 0.68. SHAP analysis shows experience and process metrics dominate, while sentiment features carry balanced, moderate importance.
We recommend that heavily loaded developers proactively communicate in direct and courteous ways to help build healthy communities, and that maintainers use the forecasting model to guide organizational restructuring.