Pseudo-User Behavioral Modeling for Software Repository Analytics in Anonymized Environments

Software repository analytics plays a critical role in improving software maintenance, project monitoring, and issue management; however, increasing privacy regulations and anonymization practices remove explicit user identifiers from issue-tracking repositories, limiting traditional analytics methods that rely on historical user information and behavioral traces. In this paper, we propose a privacy-preserving pseudo-user behavioral modeling framework for software repository analytics in anonymized environments. Our framework introduces a parallelized architecture that reconstructs latent behavioral groups, referred to as pseudo-users, using only non-personal signals extracted from issue reports. The proposed pipeline integrates BERT for semantic text representation, HDBSCAN for density-based pseudo-user discovery, temporal rhythm modeling, metadata profiling, and Latent Dirichlet Allocation (LDA) for topic extraction and behavioral characterization. Experimental evaluation on anonymized Eclipse and Mozilla datasets demonstrates that pseudo-user representations significantly enhance multiple downstream analytics tasks, including workload estimation, issue prioritization, repository trend forecasting, and behavioral anomaly detection. The proposed framework achieves a Macro-F۱ score of ۰.۸۷ and an accuracy of ۰.۸۹, consistently outperforming metadata-only, semantic-only, and conventional clustering baselines by ۶–۱۸% across evaluation metrics. In addition, parallel feature extraction and clustering reduce computational overhead and improve scalability for large-scale repositories. Comparative results establish our framework as a state-of-the-art privacy-preserving solution for behavioral analytics in anonymized software ecosystems,

کلیدواژه ها:

Index Terms—Duplicate Bug Report Detection ، Anonymized Environments ، Multi-Task Learning ، BERT