Cross-project defect prediction (CPDP) trains a defect-prediction model on data from other, completed software projects (source projects) and applies it to a new project that lacks defect data. The CFPS method (Sun et al., 2021) showed that choosing the right source projects is a distinct, consequential decision — one best treated as a recommendation problem solved with user-based collaborative filtering over project similarity and measured cross-project applicability — but it assumed source and target are described by the same software metrics, was validated on only 14 homogeneous PROMISE/Jureczko projects, and stopped at producing a ranked source list without prescribing how the recommended sources should be combined into a final training set. Later literature (2021–2026) treats the three directions the CFPS authors left open as follows: source selection continues to be re-engineered (2023–2025), heterogeneous and multi-source CPDP models proliferate, and a seven-way empirical comparison of multi-source CPDP models found no model superior on all performance measures, with new aggregation proposals still appearing in 2026. The surviving open problem is the coherent end-to-end objective the CFPS limitations point to: recommend suitable source projects when metric sets differ between source and target, and couple that recommendation with a principled aggregation/weighting and transfer strategy whose downstream defect-prediction performance is explicitly evaluated, validated on broader corpora.

Background

Software defect prediction (SDP) estimates, from code properties measured before testing, which modules of a software system are likely to contain defects, so that limited testing and review effort can be aimed at the riskiest modules. The measurable code properties used as model features are software metrics — for example counts of lines of code, or the Chidamber–Kemerer object-oriented metrics such as coupling between objects — computed per module (typically per class or file). A model trained on the project's own labeled history is called within-project defect prediction (WPDP); it fails for new projects, which have no defect-labeled history. Cross-project defect prediction (CPDP), studied since Zimmermann et al. (2009) showed that naive cross-project transfer often fails and that the choice of source project is decisive, instead trains on data from other, completed projects. Public corpora such as the PROMISE repository and the Jureczko collection of Java projects supply labeled metric tables for many open-source projects.

Research initially concentrated on instance selection — filtering individual training instances from a fixed source. Herbold (2013) instead ranked whole candidate source projects by distributional similarity to the target (EucPS: Euclidean distance between vectors of per-metric means and standard deviations) and trained on the most similar ones. The CFPS paper (Sun, Li, Sun, and He, Applied Soft Computing 99:106940, 2021) refuted the assumption that distributional similarity implies usefulness: across 14 PROMISE projects and 5 classifiers, the single most similar project was among the top-3 truly applicable sources in only 10 of 70 combinations (by AUC, the area under the receiver-operating-characteristic curve, which measures detection quality across classification thresholds). CFPS instead adapts user-based collaborative filtering — the classic recommender technique that suggests items to a user by weighting the preferences of similar users — treating the target project as a new user, candidate source projects as items, and measured cross-project prediction performance among historical projects (their pairwise 'applicability') as ratings. CFPS produced better-ranked source lists than EucPS on retrieval measures — F@N (the fraction of the top-N recommended sources that are truly applicable), MAP (mean average precision over the ranking), and MRR (mean reciprocal rank of the first applicable source) — and showed recommendation quality varies with the downstream classifier.

A parallel line, heterogeneous defect prediction (HDP; Nam and Kim 2015; Nam, Fu, Kim, Menzies, and Tan, IEEE TSE 2018), addresses learning across projects whose metric sets differ (e.g., CK metrics in the source, process metrics in the target) by matching the statistical compatibility of features rather than requiring identical feature sets. More recent multi-source CPDP (MSCPDP) work transfers from several sources at once; Liu, Li, Zou, and Tong (APSEC 2022) empirically compared seven MSCPDP models on 20 datasets using the Scott–Knott ESD test (a multiple-comparison procedure that clusters alternative methods into statistically indistinguishable groups) and found no model superior on all performance measures (the best model led on four of five measures, including AUC, but not on F1). CFPS itself acknowledges three limitations that motivate this problem: its evaluation covered only 14 projects sharing one metric set; it cannot recommend sources when the target and source metric sets differ; and it delivers a ranked list without prescribing how to aggregate multiple recommended sources (weighting instances, choosing a transfer-learning or domain-adaptation strategy — a method that aligns the source and target data distributions so that a model trained on one transfers well to the other) so that the recommendation translates into the best achievable downstream prediction. 'Negative transfer' denotes the empirically observed phenomenon that adding unsuitable source data can make prediction worse than using no transfer at all.

Problem Statement

The problem, as acknowledged in the CFPS study (Collaborative Filtering based source Projects Selection; Sun et al., 2021) and left standing by the literature since, is to turn source-project selection for cross-project defect prediction into an end-to-end procedure that works outside CFPS's original homogeneous, 14-project setting. Concretely: (1) given a target project whose modules are described by a metric set different from that of the available candidate source projects, devise a method to recommend or select suitable source projects — measured by recommendation quality (e.g., precision/recall-style measures over applicable sources, mean average precision (MAP), or mean reciprocal rank (MRR)) as well as downstream prediction performance — where 'applicability' (informally, the source yielding good defect-prediction performance on the target) must itself be given an explicit, justified operational definition (performance measure, acceptance criterion, and reference baseline) as part of the method, because defining and measuring it when source and target metric spaces differ is part of the open problem; (2) prescribe and evaluate how multiple recommended sources are actually exploited downstream — how to aggregate them into a training set, weight instances or sources, and choose the accompanying transfer-learning or domain-adaptation strategy — so that the recommendation is directly coupled to, and evaluated by, the resulting prediction performance on the target (e.g., AUC, the F-measure (the harmonic mean of precision and recall), or the Matthews correlation coefficient (MCC)); and (3) validate both on a corpus scope broader than the 14 homogeneous projects from the PROMISE repository and the Jureczko collection used for CFPS — at least two public corpora spanning at least two distinct metric sets and at least 20 projects in total — under a protocol that applies statistical testing to the downstream prediction performance of the full recommendation-and-aggregation pipeline, measured by performance measures such as AUC, the F-measure, or MCC, against relevant baselines: whenever source and target metric sets differ, against an established heterogeneous defect prediction approach; and, in the homogeneous-metric setting, against CFPS's similarity-and-applicability recommendation and EucPS (Herbold's distributional-similarity source selection), a setting that lies outside the heterogeneous scope of the contribution.

References

  1. Zhongbin Sun, Junqi Li, Heli Sun, Liang He. CFPS: Collaborative filtering based source projects selection for cross-project defect prediction. Applied Soft Computing 99:106940 (2021, available online 2020). DOI: 10.1016/j.asoc.2020.106940. https://doi.org/10.1016/j.asoc.2020.106940
  2. Thomas Zimmermann, Nachiappan Nagappan, Harald Gall, Emanuel Giger, Brendan Murphy. Cross-project defect prediction: a large scale experiment on data vs. domain vs. process. In Proceedings of the 7th Joint Meeting of the European Software Engineering Conference and the ACM SIGSOFT Symposium on the Foundations of Software Engineering (ESEC/FSE 2009). DOI: 10.1145/1595696.1595713. https://doi.org/10.1145/1595696.1595713
  3. Steffen Herbold. Training data selection for cross-project defect prediction. In Proceedings of the 9th International Conference on Predictive Models in Software Engineering (PROMISE 2013), pages 1–10. DOI: 10.1145/2499393.2499395. https://doi.org/10.1145/2499393.2499395
  4. Jaechang Nam, Sunghun Kim. Heterogeneous defect prediction. In Proceedings of the 2015 10th Joint Meeting on Foundations of Software Engineering (ESEC/FSE 2015). DOI: 10.1145/2786805.2786814. https://doi.org/10.1145/2786805.2786814
  5. Jaechang Nam, Wei Fu, Sunghun Kim, Tim Menzies, Lin Tan. Heterogeneous defect prediction. IEEE Transactions on Software Engineering 44(9) (2018). DOI: 10.1109/TSE.2017.2720603. https://doi.org/10.1109/TSE.2017.2720603
  6. Jie Wu, Yingbo Wu, Nan Niu, Min Zhou. MHCPDP: multi-source heterogeneous cross-project defect prediction via multi-source transfer learning and autoencoder. Software Quality Journal 29:405–430 (2021). DOI: 10.1007/s11219-021-09553-2. https://doi.org/10.1007/s11219-021-09553-2
  7. Xuanying Liu, Zonghao Li, Jiaqi Zou, Haonan Tong. An Empirical Study on Multi-Source Cross-Project Defect Prediction Models. In 2022 29th Asia-Pacific Software Engineering Conference (APSEC 2022), pages 318–327. DOI: 10.1109/APSEC57359.2022.00044. https://doi.org/10.1109/APSEC57359.2022.00044
  8. Jiaqi Zou, Zonghao Li, Xuanying Liu, Haonan Tong. MSCPDPLab: A MATLAB toolbox for transfer learning based multi-source cross-project defect prediction. SoftwareX 21:101286 (2023). DOI: 10.1016/j.softx.2022.101286. https://doi.org/10.1016/j.softx.2022.101286
  9. Shailza Kanwar, Lalit Kumar Awasthi, Vivek Shrivastava. Candidate project selection in cross project defect prediction using hybrid method. Expert Systems with Applications 218:119625 (2023). DOI: 10.1016/j.eswa.2023.119625. https://doi.org/10.1016/j.eswa.2023.119625
  10. Mohit Patil, Manjubala Bisi, Pravali Manchala. Source Project Selection for Cross-Project Software Defect Prediction using Clustering Approach. In 2023 IEEE 20th India Council International Conference (INDICON 2023). DOI: 10.1109/INDICON59947.2023.10440956. https://doi.org/10.1109/INDICON59947.2023.10440956
  11. Pravali Manchala, Manjubala Bisi. A study on cross-project fault prediction through resampling and feature reduction along with source projects selection. Automated Software Engineering 31 (2024). DOI: 10.1007/s10515-024-00465-6. https://doi.org/10.1007/s10515-024-00465-6
  12. Pravali Manchala, Manjubala Bisi. A novel source project and optimized training data selection approach for cross-project fault prediction. The Journal of Supercomputing 81 (2024). DOI: 10.1007/s11227-024-06750-1. https://doi.org/10.1007/s11227-024-06750-1
  13. Nazgol Nikravesh, Mohammad Reza Keyvanpour. MCOT-KB: Multi-source cross-project defect prediction with optimal transport domain adaptation and KMMBagging. Software Quality Journal 34(1) (2026). DOI: 10.1007/s11219-025-09739-y. https://doi.org/10.1007/s11219-025-09739-y
  14. Lifang Chen, Kexin Cao, Sipeng Zhang, Renzhe Zhao, Qi Dai. Multi-source heterogeneous software defect prediction via SVM optimized by filtered mutation projection ivy algorithm. The Journal of Supercomputing 82(10) (2026). DOI: 10.1007/s11227-026-08688-y. https://doi.org/10.1007/s11227-026-08688-y
  15. Rizal Broer Bahaweres, Nabila Fatia Kasmizar. Comparion of Source-Selection Between KNN-TDS Method and Median Z-Score in Multi-Source Cross-Project Defect Prediction Using Neural Network. In 2025 13th International Conference on Cyber and IT Service Management (CITSM 2025). DOI: 10.1109/CITSM67730.2025.11291213. https://doi.org/10.1109/CITSM67730.2025.11291213