Wednesday, November 18, 2026
Science
No Result
View All Result
  • Login
  • HOME
  • SCIENCE NEWS
  • CONTACT US
  • HOME
  • SCIENCE NEWS
  • CONTACT US
No Result
View All Result
Scienmag
No Result
View All Result
Home Science News Technology and Engineering

Outlier-Conscious Oversampling Method Improves Cross-Project Software Defect Prediction Performance Across Benchmarks

November 18, 2026
in Technology and Engineering
Denise Maddox
By Denise Maddox Scienmag Editorial Profile - Mechanical Engineering
Reading Time: 7 mins read
0
Outlier-Conscious Oversampling Method Improves Cross-Project Software Defect Prediction Performance Across Benchmarks

Outlier-Conscious Oversampling Method Improves Cross-Project Software Defect Prediction Performance Across Benchmarks

65
SHARES
587
VIEWS
Share on FacebookShare on Twitter
ADVERTISEMENT

Software engineering researchers have demonstrated that filtering atypical observations during domain-adaptive minority oversampling significantly improves the accuracy of cross-project defect prediction models across extensive benchmark datasets. In a newly published investigation, a research team introduced Outlier-Conscious Transfer-Oriented Minority Oversampling, designated as OTOMO, and evaluated its predictive capabilities across sixty-six distinct source-to-target project pairs drawn from four widely recognized software defect archives: NASA, RELINK, SOFTLAB, and AEEEM. When paired with the parameter-free domain adaptation method known as Correlation Alignment, or CORAL, the newly formulated approach achieved the highest overall average performance across four standard evaluation metrics using both ensemble-based Random Forest and linear Logistic Regression classifiers. Specifically, across all tested project transfer scenarios, the framework recorded an overall average area under the receiver operating characteristic curve of 73.83 percent with Random Forest and 68.00 percent with Logistic Regression. Furthermore, the method registered average G-mean scores of 60.70 percent and 60.12 percent, average F1-scores of 38.40 percent and 36.10 percent, and average Matthews Correlation Coefficients of 27.32 percent and 23.12 percent for the two classifiers, respectively, surpassing conventional oversampling baselines and distribution adaptation methods that do not jointly account for target-domain distributions and outlier contamination.

Software defect prediction functions as an indispensable mechanism in contemporary software quality assurance by pinpointing program components that exhibit an elevated likelihood of containing faults prior to operational release. Because exhaustive software testing is exceptionally demanding in terms of manual effort, developer hours, and financial expenditure, accurately forecasting fault-prone components allows development teams to allocate limited verification resources with greater efficiency. Empirical defect distribution studies have consistently revealed that the substantial majority of software defects tend to cluster within a relatively small fraction of software components. Traditionally, defect prediction frameworks have operated under a within-project defect prediction paradigm, training predictive models on historical fault data gathered from earlier iterations or modules of the identical software repository. However, newly developed software projects and recently launched repositories invariably suffer from an absence of historical execution metrics and verified defect annotations. Labeling modules manually requires intensive domain expertise and prohibitive timelines. To resolve this cold-start dilemma, cross-project defect prediction utilizes labeled metric data harvested from external, mature source projects to construct predictive classifiers capable of identifying defect-prone components in unannotated target projects.

Despite its conceptual utility, transferring defect prediction models across disparate software codebases introduces two severe mathematical and structural bottlenecks: pronounced class imbalance and substantial feature distribution discrepancy. Software defect datasets are inherently imbalanced because non-defective components outnumber defective instances by a wide margin, often causing conventional supervised learning algorithms to skew heavily in favor of the majority class while failing to detect defect-prone modules. Simultaneously, metric values vary substantially across systems due to differences in software architecture, team conventions, development methodologies, and project scale, resulting in acute distribution shifts between source and target datasets. To alleviate feature discrepancies, researchers previously introduced unsupervised domain adaptation techniques such as Transfer Component Analysis, which utilizes Maximum Mean Discrepancy to minimize marginal distribution differences in a reproducing kernel Hilbert space, as well as normalized extensions like TCA+. Subsequent methodologies such as Joint Distribution Adaptation and Balanced Distribution Adaptation incorporated conditional distributions alongside pseudo-labels, though their operational efficiency is constrained by extensive hyperparameter tuning requirements. In contrast, Correlation Alignment emerged as a computationally efficient, parameter-free alternative that aligns the covariance matrices of source and target data to align second-order statistical profiles.

Although domain adaptation strategies mitigate overall distribution mismatches between software repositories, they do not resolve the pervasive challenge of class imbalance within training sets. Conversely, conventional oversampling algorithms, including the Synthetic Minority Over-sampling Technique, Adaptive Synthetic Sampling, and Determinantal Point Process sampling, balance training distributions by generating artificial minority instances along interpolation lines or prioritizing diverse sample subsets within the feature space. Nevertheless, these classic techniques were originally designed for isolated, single-project settings and construct synthetic instances relying exclusively on source-domain neighborhood structures, without considering whether newly created instances align with the feature space of the target project. To bridge this divide, prior literature introduced Transfer-oriented Minority Oversampling, known as TOMO, which guides synthetic instance generation using target-domain proximity. However, the experimental evidence highlights that both conventional oversampling algorithms and transfer-oriented strategies remain vulnerable to distortion when training sets contain outliers. Because distance-based interpolation relies heavily on localized sample neighborhoods, atypical observations or noisy boundary samples can mislead synthetic generation mechanisms, creating unrepresentative synthetic instances in ambiguous regions of the feature landscape.

The OTOMO algorithm resolves these compounded limitations through a structured, three-stage generation procedure that integrates unsupervised target profiling, rigorous outlier pruning, and hybrid neighborhood construction. In the initial phase, because ground-truth defect annotations are unavailable for the target repository, the framework applies unsupervised k-means clustering with two centroids to partition the target feature space, assigning the smaller partition as the target pseudo-minority cluster. To prevent atypical components from corrupting subsequent sampling phases, the algorithm computes the squared Mahalanobis distance for each instance in the target pseudo-minority cluster and discards observations exceeding a 99.5 percent confidence threshold derived from the chi-square distribution quantile. In the second phase, the framework calculates pairwise Euclidean distances among source minority instances alongside their respective distances to the cleaned target pseudo-minority centroid, merging these two spatial components into a hybrid distance matrix controlled by a trade-off parameter, denoted as gamma, bounded between zero and one. In the third phase, synthetic minority samples are synthesized using the hybrid distance structure to identify the seven nearest neighbors for each source minority instance, ensuring that newly generated observations simultaneously reflect the underlying source minority geometry and the filtered distribution of the target domain.

Following minority class oversampling, the OTOMO framework completes cross-project adaptation by integrating parameter-free Correlation Alignment to synchronize the balanced source metric space with the target feature distribution. Software metrics across both repositories are first standardized using standard Z-score normalization to enforce zero mean and unit variance across individual features. Because covariance matrix estimation can encounter numerical instability when projects contain limited instance counts relative to dimensionality, the diagonal elements of both the balanced source covariance matrix and the target covariance matrix are regularized by adding an identity adjustment factor to ensure full rank invertibility. The alignment procedure subsequently executes a two-stage mathematical transformation comprising feature whitening followed by re-coloring, thereby directly aligning the second-order statistical covariance of the augmented source training set with that of the target system. Unlike earlier distribution adaptation methods that require iterative optimization or complex parameter searches across individual repository pairs, combining OTOMO with Correlation Alignment yields an end-to-end training pipeline that stabilizes minority sample transfer without requiring manual trade-off tuning or target label exposure.

Within each project group, every repository was systematically deployed as a source project to predict defects across each remaining peer project, generating sixty-six distinct cross-project transfer configurations, comprising twenty pairs for NASA, six for RELINK, twenty for SOFTLAB, and twenty for AEEEM. Statistical profiling confirmed that outlier observations were prevalent across all evaluated repositories, ranging from 9.33 percent in NASA project PC4 to 91.67 percent in SOFTLAB project AR5, where an extremely constrained sample size of thirty-six instances across twenty-nine metrics produced pronounced covariance instability. Predictive experiments were repeated across thirty independent random seeds to mitigate stochastic variation, evaluating OTOMO against four baseline oversampling approaches and two standalone domain adaptation techniques using both Random Forest and Logistic Regression.

Detailed comparative performance tabulations demonstrated that OTOMO secured the highest aggregate averages across all four evaluated performance metrics and both classification architectures. When evaluated using Random Forest, OTOMO achieved an overall average ROC-AUC of 73.83 percent, exceeding DPP at 73.24 percent, SMOTE at 72.87 percent, ADASYN at 72.17 percent, standalone CORAL at 72.13 percent, TCA+ at 69.05 percent, and TOMO at 63.29 percent. In head-to-head win-tie-loss tallies across the sixty-six project pairs under Random Forest, OTOMO recorded thirty-six wins, two ties, and twenty-eight losses against DPP, thirty-eight wins against SMOTE, forty-two wins against ADASYN, and fifty-six wins against TOMO. For threshold-dependent metrics under Random Forest, OTOMO attained a mean G-mean of 60.70 percent, outperforming ADASYN at 59.76 percent and SMOTE at 59.68 percent while establishing substantial margins over DPP at 50.98 percent, CORAL at 38.55 percent, and TCA+ at 26.89 percent.

To confirm that the observed empirical improvements were not artifacts of random sampling, the authors conducted non-parametric statistical hypothesis testing utilizing the Friedman rank test with the Iman-Davenport extension, evaluated at a significance threshold of alpha equal to 0.05. Across all combinations of classifiers and evaluation metrics, the calculated Iman-Davenport statistics ranged from 12.53 to 56.24, significantly surpassing the critical threshold value of 2.12 and conclusively rejecting the null hypothesis of equivalent performance across methods. Subsequent post-hoc evaluation utilizing the Nemenyi test with a critical difference threshold of approximately 1.109 demonstrated that OTOMO secured the premier average rank in seven of the eight experimental metric-classifier configurations, yielding statistically significant improvements over comparative baselines in thirty of forty-eight pairwise comparisons while never performing significantly worse. Evaluating results across individual dataset groups revealed that OTOMO secured the top group average in twenty-three of thirty-two potential evaluations, dominating all eight metric configurations within NASA, seven within AEEEM, and six within SOFTLAB, while ranking among the top three methods across every evaluated group scenario.

The experimental analysis also delineated key operational boundaries, localized performance variations, and parameters influencing model stability across heterogeneous environments. In the RELINK group, which exhibited the lowest clustering purity ranging between 58.25 percent and 73.21 percent across its three projects, OTOMO achieved the highest average in only two of eight configurations, as project Apache possessed an approximately balanced defect ratio of 50.52 percent that precluded oversampling, while TCA+ achieved superior transfer results under Logistic Regression. Sensitivity analyses across the trade-off parameter gamma confirmed that prioritizing target-domain guidance through smaller values between 0.0 and 0.2 yielded optimal predictive stability across fifteen of sixteen evaluated response curves, corroborating the selection of a fixed gamma value of 0.1 to avoid target label leakage. The authors noted that while OTOMO established superior aggregate reliability, isolated transfer losses occurred when target minority estimation degraded or when outlier pruning inadvertently removed informative minority instances. Consequently, future research initiatives aim to investigate advanced unsupervised clustering frameworks, automated hyperparameter selection mechanisms, and complementary feature extraction methodologies to further enhance cross-project defect prediction across industrial software ecosystems.

Subject of Research: Machine learning domain adaptation and minority oversampling for cross-project software defect prediction

Article Title: OTOMO: Outlier-Conscious Transfer-Oriented Minority Oversampling for cross-project defect prediction

Article References: Akhter, S., Rana, M. J., Farid, F. A., Uddin, J., Abdul Karim, H. B., & Kibria, M. G. (2026). OTOMO: Outlier-Conscious Transfer-Oriented Minority Oversampling for cross-project defect prediction. Frontiers in Computer Science, 8, Article 1915086. https://doi.org/10.3389/fcomp.2026.1915086

Image Credits: AI Generated

DOI: 10.3389/fcomp.2026.1915086

Keywords: cross-project defect prediction, software engineering, minority oversampling, outlier detection, domain adaptation, correlation alignment, class imbalance, machine learning, random forest, logistic regression, NASA dataset, RELINK dataset, SOFTLAB dataset, AEEEM dataset, Mahalanobis distance, k-means clustering, synthetic sampling

Cite Scienmag News

Denise Maddox. (November 18, 2026). Outlier-Conscious Oversampling Method Improves Cross-Project Software Defect Prediction Performance Across Benchmarks. Scienmag. https://scienmag.com/outlier-conscious-oversampling-method-improves-cross-project-software-defect-prediction-performance-across-benchmarks/

Denise Maddox. "Outlier-Conscious Oversampling Method Improves Cross-Project Software Defect Prediction Performance Across Benchmarks." Scienmag, 18 November 2026, https://scienmag.com/outlier-conscious-oversampling-method-improves-cross-project-software-defect-prediction-performance-across-benchmarks/. Accessed 18 November 2026.

Denise Maddox. "Outlier-Conscious Oversampling Method Improves Cross-Project Software Defect Prediction Performance Across Benchmarks." Scienmag. November 18, 2026. https://scienmag.com/outlier-conscious-oversampling-method-improves-cross-project-software-defect-prediction-performance-across-benchmarks/

Tags: AEEEM datasetclass imbalancecorrelation alignmentcross-project defect predictiondomain adaptationK-means clusteringlogistic regressionMachine learningMahalanobis distanceminority oversamplingNASA datasetoutlier detectionRandom ForestRELINK datasetSOFTLAB datasetsoftware engineeringsynthetic sampling
Share26Tweet16
Previous Post

Laboratory Tests Link Olivine Dissolution to Seawater Alkalinity Despite Tracer Clumping

Next Post

Conserved Trehalose Pathway Aids Acidophilic Archaeon Survival Under Heat and Salinity

Related Posts

Meta-Learning Accelerates Microwave Hyperthermia Antenna Optimization in Proof-of-Concept Study
Medicine

Meta-Learning Accelerates Microwave Hyperthermia Antenna Optimization in Proof-of-Concept Study

November 18, 2026
Rethinking Inertial Terms in Halpern-Type Iterations for Strong Convergence
Technology and Engineering

Rethinking Inertial Terms in Halpern-Type Iterations for Strong Convergence

November 18, 2026
Reliability-Aware Query Modulation Improves Accuracy in Transformer-Based Multi-Object Tracking
Technology and Engineering

Reliability-Aware Query Modulation Improves Accuracy in Transformer-Based Multi-Object Tracking

November 18, 2026
GATOC-ERC Framework Combines Conversational Graphs and Language Models for Emotion Recognition
Technology and Engineering

GATOC-ERC Framework Combines Conversational Graphs and Language Models for Emotion Recognition

November 18, 2026
MulDataFrame Accelerates Metadata Editing Over 12,000-Fold While Reducing Memory Footprint
Technology and Engineering

MulDataFrame Accelerates Metadata Editing Over 12,000-Fold While Reducing Memory Footprint

November 18, 2026
Co-Milling Serpentine With Calcite Boosts Portland Cement Strength By Regulating Magnesium
Technology and Engineering

Co-Milling Serpentine With Calcite Boosts Portland Cement Strength By Regulating Magnesium

November 18, 2026
Next Post
Conserved Trehalose Pathway Aids Acidophilic Archaeon Survival Under Heat and Salinity

Conserved Trehalose Pathway Aids Acidophilic Archaeon Survival Under Heat and Salinity

  • Mothers who receive childcare support from maternal grandparents show more optimized

    Mothers who receive childcare support from maternal grandparents show more parental warmth, finds NTU Singapore study

    27656 shares
    Share 11059 Tweet 6912
  • University of Seville Breaks 120-Year-Old Mystery, Revises a Key Einstein Concept

    1061 shares
    Share 424 Tweet 265
  • Bee body mass, pathogens and local climate influence heat tolerance

    682 shares
    Share 273 Tweet 171
  • Researchers record first-ever images and data of a shark experiencing a boat strike

    546 shares
    Share 218 Tweet 137
  • Groundbreaking Clinical Trial Reveals Lubiprostone Enhances Kidney Function

    531 shares
    Share 212 Tweet 133
Science

Embark on a thrilling journey of discovery with Scienmag.com—your ultimate source for cutting-edge breakthroughs. Immerse yourself in a world where curiosity knows no limits and tomorrow’s possibilities become today’s reality!

RECENT NEWS

  • Combined Malnutrition and Poor Quality of Life Elevate Digestive Cancer Mortality
  • Advances in Humanized Mouse Models Enhance Immunological Disease Research and Drug Discovery
  • Meta-Learning Accelerates Microwave Hyperthermia Antenna Optimization in Proof-of-Concept Study
  • Beta-Glucans Show Emerging Cervical Cancer Potential Despite Major Translational Gaps

Categories

  • Agriculture
  • Anthropology
  • Archaeology
  • Athmospheric
  • Biology
  • Biotechnology
  • Blog
  • Bussines
  • Cancer
  • Chemistry
  • Climate
  • Earth Science
  • Editorial Policy
  • Marine
  • Mathematics
  • Medicine
  • Pediatry
  • Policy
  • Psychology & Psychiatry
  • Science Education
  • Science News
  • Social Science
  • Space
  • Technology and Engineering

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 5,151 other subscribers

© 2026 Scienmag - Science Magazine

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • HOME
  • SCIENCE NEWS
  • CONTACT US

© 2026 Scienmag - Science Magazine

Discover more from Science

Subscribe now to keep reading and get access to the full archive.

Continue reading