A hybrid machine learning framework combining clustering, generative synthetic oversampling, and stacked ensemble modeling has demonstrated improved accuracy, recall, and discriminative performance in forecasting employee attrition, according to a study published in the journal Human-Centric Intelligent Systems. Researchers Akhilesh Kumar Singh and Koushlendra Kumar Singh designed the integrated pipeline to resolve two major computational hurdles in workforce analytics: the presence of heterogeneous employee segments and severe class imbalance, in which departing employees represent a small minority of the observed workforce. Modern enterprise management depends on reliable attrition forecasts to support strategic planning, operational effectiveness, and organizational stability, yet conventional machine learning models frequently struggle with skewed data distributions, resulting in poor sensitivity to voluntary resignations.
To evaluate the methodology, the researchers utilized a benchmark dataset comprising 14,999 employee records across ten attributes, sourced from the Kaggle platform and evaluated alongside established IBM HR Analytics standards. The feature set incorporated numerical indicators such as employee satisfaction level, evaluation scores, the number of assigned projects, average monthly working hours, and tenure in years, alongside binary variables for workplace accidents and promotions in the prior five years. Categorical variables captured organizational department and salary tiers, while turnover status served as the primary binary classification target. Exploratory analysis confirmed that the underlying dataset exhibited no missing values, establishing a consistent foundation for modeling multivariate employee behaviors.
Initial distribution analysis revealed distinct operational and behavioral divergences between retained workers and departing personnel. Kernel density estimation showed that employees who left the organization exhibited persistently low satisfaction scores, with distinct density peaks below 0.4, whereas retained employees demonstrated peak density between 0.6 and 0.9. Evaluation scores for departing employees displayed a bimodal distribution with peaks near 0.5 and 0.85, indicating heightened turnover risks among both under-evaluated staff and heavily evaluated individuals. Project allocations further highlighted divergent dynamics: employees assigned to only two or three projects, as well as those managing excessive commitments, exhibited elevated attrition, whereas personnel working on four to five projects remained predominantly stable.
Workplace hours, organizational tenure, and compensation factors exhibited similarly clear associations with employee separation patterns. Departing employees displayed bimodal peaks in monthly commitments, congregating around 150 hours and 275 hours, in contrast to the stable 200-hour midpoint observed among retained personnel. Resignations peaked between three and four years of company tenure before declining sharply, corroborating organizational trends where turnover concentrates in early-to-mid career phases. Across categorical dimensions, separation rates were highest within low- and medium-salary brackets, while employees who had received a promotion within the preceding five years experienced substantially lower turnover than unpromoted counterparts, emphasizing the predictive importance of internal mobility and compensation.
The methodological framework developed by the authors addresses data complexity through four sequential phases: preprocessing, mixed-type employee segmentation, synthetic minority oversampling, and ensemble classification. In the preprocessing phase, categorical features were encoded, numerical attributes normalized, and pairs exhibiting high pairwise correlations exceeding an absolute value of 0.85 were eliminated to mitigate multicollinearity. Following data partitioning into training and testing subsets, the researchers deployed K-Prototypes clustering to partition the workforce into homogeneous subgroups. Because human resources datasets incorporate both discrete variables like job department and continuous measures such as working hours, K-Prototypes minimizes an objective function penalizing both numerical variance and categorical dissimilarity via an indicator function.
The resulting cluster prototypes served to uncover localized behavioral patterns that global classification models typically miss. Once samples were assigned to their nearest cluster centroid, the corresponding cluster label was appended directly to the feature vector representation as an explicit structural feature. The authors observed that class imbalance was not uniform across the dataset but varied widely by cluster: while certain clusters were heavily dominated by retained employees, others contained substantial concentrations of resignations. This finding demonstrated that class imbalance was fundamentally cluster-dependent, requiring tailored data augmentation rather than uniform global resampling across the entire population.
To rectify the local minority underrepresentation without distorting complex feature interactions, the researchers integrated Conditional Tabular Generative Adversarial Networks into the modeling pipeline. Unlike standard linear interpolation techniques such as SMOTE, which often fail to capture complex categorical dependencies and high-cardinality distributions, the conditional generative adversarial network utilizes mode-specific normalization and conditional sampling to generate high-fidelity synthetic minority instances. Augmentation was applied selectively to clusters demonstrating acute class deficits, equilibrating label distributions while leaving adequately populated clusters unchanged. Subsequent fairness and distribution audits verified that synthetic sample generation preserved joint attribute correlations without skewing cluster boundaries or test set distributions.
Classification was executed through a stacked ensemble architecture incorporating diverse base learners, including artificial neural networks, extreme gradient boosting, and CatBoost models. Base learners were trained across cross-validation folds to generate out-of-fold prediction vectors, preventing overfitting and information leakage before being supplied to a meta-classifier. When evaluated across the segmented clusters, the proposed model achieved overall classification accuracy between 0.9857 and 0.9926, consistently outperforming standalone neural networks and traditional machine learning baselines such as Support Vector Machines, Naive Bayes, and standard Logistic Regression. The ensemble achieved an overall area under the receiver operating characteristic curve of 0.9928 across the combined cohort, with individual clusters reaching scores up to 0.9972.
Comparative analyses revealed that traditional models trained on non-clustered data exhibited severe performance bottlenecks, particularly regarding minority-class recall. While standard logistic regression and support vector machines attained respectable nominal accuracy, their recall scores reached only 0.23 and 0.15, respectively, indicating a pronounced inability to detect genuine attrition cases. By contrast, the hybrid model maintained balanced sensitivity, posting recall values near 0.97 across multiple cluster partitions. Bootstrap confidence interval analyses comparing cluster subsets confirmed that while classification accuracy remained consistently elevated across segments, variations in sensitivity and harmonic metric scores reflected genuine structural differences in underlying workforce risk profiles.
The authors concluded that integrating unsupervised segmentation, generative oversampling, and stacked ensemble modeling provides an effective and interpretable framework for human resources decision-making and retention planning. Nonetheless, the researchers highlighted several technical boundaries and computational considerations inherent to the workflow. The architecture relies on specific combinations of base models and meta-learners, meaning alternative configurations or diverse organizational datasets could yield varying empirical outcomes. Furthermore, the extensive k-fold cross-validation, iterative generative model fitting, and fine-tuning procedures substantially increase computational resource demands, presenting operational trade-offs when scaling the methodology to exceptionally large corporate datasets.
To rigorously evaluate the model’s ability to distinguish between retention and turnover conditions, the researchers utilized a comprehensive set of performance metrics that extended beyond simple accuracy. These included precision, recall, the F1-score, the geometric mean, and the area under the receiver operating characteristic curve. By examining the binary confusion matrix, which tallies true positives, true negatives, false positives, and false negatives, the team gained a more granular understanding of classifier efficacy across varied label distributions. This multi-metric approach was particularly critical in scenarios where the majority class overwhelmed the minority, as it allowed for a balanced assessment of how effectively the system identified genuine attrition instances without triggering excessive false alarms.
A core element of the methodology was the use of consensus feature ranking, which provided a clearer view of the specific drivers behind employee decisions. By aggregating feature importance across multiple methods and breaking ties using mutual information patterns, the researchers identified key attributes that disproportionately correlate with attrition. The analysis confirmed that financial rewards and career advancement pathways were the most influential predictors, outweighing factors such as workplace accidents. Specifically, employees earning low or medium salaries, coupled with a lack of promotion over a five-year period, exhibited the highest rates of departure. These insights emphasize that organizational interventions aimed at adjusting compensation structures and refining internal promotion policies could likely yield measurable improvements in retention metrics.
Fairness and distributional integrity were also central to the validation of the synthetic data generation process. Before and after employing the conditional tabular generative adversarial network, the researchers conducted systematic audits to ensure that the augmented data did not inadvertently introduce disparities or skew the inherent feature correlations. By applying rule-based constraint validation and distributional similarity tests, the team confirmed that the synthetic minority samples successfully filled local class deficits without distorting the cluster boundaries. This rigorous validation ensures that the synthetic samples remain representative of realistic employee behaviors rather than artifacts of the modeling pipeline.
The study also utilized bootstrap confidence intervals to provide a statistical perspective on the differences in classification performance between employee subgroups. When comparing specific segments, such as Cluster 0 and Cluster 1, the confidence intervals indicated stable and reliable results, with Cluster 1 showing higher sensitivity and better balanced harmonic performance scores. These findings underscore the premise that class imbalance in human resources data is often cluster-dependent, rather than a universal characteristic of the entire workforce. Consequently, the researchers’ decision to implement distribution-based resampling at the cluster level, rather than applying uniform global oversampling, was fundamental to the framework’s improved recall performance.
Ultimately, the researchers demonstrated that integrating unsupervised segmentation, generative augmentation, and ensemble learning facilitates a predictive mechanism that is both scalable and accurate. The study highlights that while individual base learners like artificial neural networks or gradient boosting may provide high performance, the synergy created by the stacking meta-learner produces superior generalization. This integrated approach allows HR professionals to move beyond basic turnover projections toward more strategic, data-driven workforce planning that accounts for the behavioral heterogeneity inherent in modern organizations.
Subject of Research: Machine learning framework for predicting employee attrition using clustering, generative oversampling, and ensemble modeling
Article Title: Mitigating Class Imbalance in Human Resources Attrition Prediction using Hybrid Generative and Cluster-Driven Ensemble Modeling
Article References: Singh, A. K., & Singh, K. K. (2026). Mitigating Class Imbalance in Human Resources Attrition Prediction using Hybrid Generative and Cluster-Driven Ensemble Modeling. Human-Centric Intelligent Systems, 6(2), 354-365. https://doi.org/10.1007/s44230-026-00146-5
Image Credits: AI Generated
DOI: 10.1007/s44230-026-00146-5
Keywords: employee attrition, machine learning, workforce analytics, K-Prototypes clustering, Conditional Tabular GAN, CTGAN, stacking ensemble, class imbalance, human resources analytics, predictive modeling, artificial neural networks, CatBoost, XGBoost, synthetic data, data augmentation, employee retention, classification metrics
Cite Scienmag News
Teresa Odom. (November 19, 2026). Hybrid Machine Learning Framework Improves Accuracy in Predicting Employee Attrition. Scienmag. https://scienmag.com/hybrid-machine-learning-framework-improves-accuracy-in-predicting-employee-attrition/
Teresa Odom. "Hybrid Machine Learning Framework Improves Accuracy in Predicting Employee Attrition." Scienmag, 19 November 2026, https://scienmag.com/hybrid-machine-learning-framework-improves-accuracy-in-predicting-employee-attrition/. Accessed 19 November 2026.
Teresa Odom. "Hybrid Machine Learning Framework Improves Accuracy in Predicting Employee Attrition." Scienmag. November 19, 2026. https://scienmag.com/hybrid-machine-learning-framework-improves-accuracy-in-predicting-employee-attrition/








