Interpretable Temporal Risk Modeling for Contributor Inactivity Prediction: A Comparative Study of Tree-Based Ensembles

Abstract

This study aims to develop an interpretable temporal risk modeling framework for predicting contributor inactivity in collaborative development environments, thereby supporting sustained participation and improving productivity. The research focuses on contributor activity data collected from a collaborative software development platform, in which participation histories are represented by temporal engagement features that capture activity recency, participation intensity, and contribution patterns over time. To model inactivity risk, several tree-based ensemble learning algorithms, including Random Forest, XGBoost, LightGBM, and a stacking ensemble, are employed and evaluated under imbalanced classification conditions. Experimental results demonstrate strong predictive performance across models, with Random Forest achieving the highest AUC of 0.9401, while XGBoost obtains the best Matthews Correlation Coefficient (0.7353). The novelty of this study lies in prioritizing structured temporal behavioral representation through normalized temporal engagement features rather than increasing model complexity, enabling more interpretable inactivity risk modeling. The findings provide practical implications for collaborative platform managers by enabling early identification of contributor disengagement, supporting sustained participation, improving productivity, and facilitating continuous product innovation.