EUVIMEDEUVIMED entdecken
Crossref

Cross‐Repository Software Defect Prediction: A Leave‐One‐Repository‐Out Empirical Study Across Five Mature Open‐Source Projects

Vijay Prasad Javvadi

Journal of Software: Evolution and Process · 2026-11 · 38 (11)

DOI: 10.1002/smr.70182 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Software: Evolution and Process, online erschienen am 16.11.2026. Verfasst von Vijay Prasad Javvadi. Die Arbeit nennt 34 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

We report a leave‐one‐repository‐out (LORO) study of file‐level software defect prediction across five mature open‐source systems—Elasticsearch, Spring Boot, Hadoop, Kafka, and Express—totalling 296,457 file instances. Classifiers are trained on the union of four repositories and evaluated on the fifth, using seven process metrics computable from any Git history; defect labels come from a six‐keyword commit‐message heuristic. Transfer holds and is not classifier‐specific: across a six‐classifier panel, every model attains cross‐repository AUC > 0.80 on every held‐out target, with Gradient Boosting best (mean AUC 0.8672) and the Random Forest reference configuration at 0.8598 (per‐target 0.8106 on Express to 0.9019 on Spring Boot, with bootstrap 95% confidence intervals). Mean cross‐repository is 0.6362 versus a within‐repository mean of 0.6834, a 0.047 absolute ( 7% relative) degradation—essentially unchanged (−0.046) when both models are scored on identical matched test rows; AUC degrades by only 0.028, evidence that score rankings transfer better than fixed‐threshold decisions. On the smallest target (Express), the cross‐repository model matches its within‐repository baseline ( 0.7005 vs. 0.6846 full‐target; parity on matched rows). The conclusions survive four pre‐registered robustness probes: capping every source repository at 10,000 instances so no repository dominates training (mean AUC −0.005); removing documentation and test files and injecting up to 20% label noise ( AUC loss); adding size and cyclomatic‐complexity code metrics (process 0.868, code‐only 0.645, combined 0.871 on a matched subset); and a strict past‐to‐future temporal protocol on re‐mined histories (mean Gradient‐Boosting AUC 0.848 0.790, all targets 0.71–0.84). We implement no domain‐adaptation algorithms; the contribution is a large, fully reproducible LORO baseline—with archived per‐instance predictions—on which subsequent adaptation work can be evaluated.

Zitieren

APA

Javvadi, V. P. (2026). Cross‐Repository Software Defect Prediction: A Leave‐One‐Repository‐Out Empirical Study Across Five Mature Open‐Source Projects. Journal of Software: Evolution and Process, 38(11). https://doi.org/10.1002/smr.70182

Vancouver

Javvadi VP. Cross‐Repository Software Defect Prediction: A Leave‐One‐Repository‐Out Empirical Study Across Five Mature Open‐Source Projects. Journal of Software: Evolution and Process. 2026;38(11). doi:10.1002/smr.70182

BibTeX
@article{Javvadi2026,
  author = {Javvadi, Vijay Prasad},
  title = {Cross‐Repository Software Defect Prediction: A Leave‐One‐Repository‐Out Empirical Study Across Five Mature Open‐Source Projects},
  journal = {Journal of Software: Evolution and Process},
  year = {2026},
  volume = {38},
  number = {11},
  doi = {10.1002/smr.70182}
}
Kostenlos starten →