A Deep Learning Model Based on Bidirectional Temporal Convolutional Network (Bi-TCN) for Predicting Employee Attrition
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsIn this paper, authors present a deep learning framework based on a Bidirectional Temporal Convolutional Network (Bi-TCN) to predict employee attrition. They use IBM HR Analytics data and apply a GAN-based data augmentation technique to create a more balanced dataset. The model achieves 92.10% accuracy on the original dataset and 96.55% accuracy on the synthetic dataset, outperforming existing machine learning and deep learning models.
In general, the paper's technical content is average but the work is of interest for larger community. The paper is also well written with good organization structure.
Some comments:
- Fig. 2 Bi-TCN text not readable.
- The experiments rely solely on one dataset (IBM HR Analytics). Since the model is trained only on IBM data, it may not generalize well to industries with different attrition patterns.
- The Generative Adversarial Network (GAN) creates synthetic employee records, but it is not clear how close to reality is this. Can you compare synthetic vs. real employee data to ensure realism?
- The IBM dataset is small (1,470 records), and deep learning models typically require larger datasets. Can you discuss on this issue?
- Bi-TCN requires significant computational resources due to its bidirectional nature and multiple convolutional layers. How long did training occur?
Author Response
Please see the attached file.
Author Response File: Author Response.docx
Reviewer 2 Report
Comments and Suggestions for AuthorsComments and Potential Improvements
-
What is the main question addressed by the research?
The study attempts to predict employee attrition using a deep learning model based on a Bidirectional Temporal Convolutional Network (Bi-TCN). However, the problem itself is not new, and many existing studies have already explored similar predictive modeling techniques. The research question lacks novelty, as numerous machine learning and deep learning models have already been applied to employee attrition prediction with comparable or even superior results. -
Is the topic original and relevant? Does it address a gap in the field?
- The topic itself is relevant but far from original.
- The study does not introduce a fundamental breakthrough in employee attrition prediction but rather repackages existing approaches.
- The use of GAN for data augmentation is not particularly innovative, as synthetic data generation techniques have been widely used in predictive modeling. The authors fail to justify why GAN-based augmentation is necessary instead of more conventional resampling methods such as SMOTE or data weighting.
- The research does not address any significant gap in the field, as previous studies have already demonstrated high accuracy in attrition prediction using traditional models such as Random Forest, XGBoost, and LSTMs.
- There is no discussion on practical applications, such as how businesses should interpret and utilize these predictions beyond generating raw accuracy metrics.
-
What does the research add to the subject area compared with other published material?
- The research does not offer substantial improvements over existing literature.
- The reported accuracy improvements (97.69%) seem unrealistic and possibly a result of overfitting, particularly due to GAN-generated synthetic data, which may not truly reflect real-world variations.
- The comparison with traditional machine learning models is superficial; the authors fail to provide statistical significance testing for their reported improvements.
- The study lacks explainability and interpretability, which are crucial for real-world HR applications. Simply reporting a high accuracy does not help HR professionals understand why employees are leaving, which is more important than just predicting whether they will leave.
-
What specific improvements should the authors consider regarding the methodology?
- The justification for using Bi-TCN is weak—the study does not demonstrate why this architecture is better suited for the problem than existing models such as LSTM, Bi-GRU, or even simple ensemble methods.
- The dataset choice is highly limited—the study is entirely based on one dataset (IBM HR Analytics), which is relatively small (only 1,470 records). There is no attempt to validate findings on different datasets, raising serious concerns about the model’s generalizability.
- GAN-based data augmentation introduces potential biases, and the authors do not provide any validation to ensure that the synthetic data does not distort the real-world distributions.
- The hyperparameter selection process is unclear—there is no explanation of how the optimal hyperparameters were chosen, suggesting possible bias in favor of the proposed model.
- Feature engineering is not well justified—certain features may have been irrelevant or redundant, yet the study does not provide any feature importance analysis to justify their inclusion.
-
Are the conclusions consistent with the evidence and arguments presented?
- The high accuracy results seem suspicious, particularly given that many previous studies using stronger models (e.g., Random Forest, CatBoost, XGBoost) achieved lower scores. This raises concerns about data leakage, overfitting, or an inflated evaluation metric.
- The authors claim that their model “outperforms existing methods”, but this statement is misleading without extensive benchmarking across multiple datasets and real-world applications.
- There is no discussion of model limitations—the authors do not address ethical concerns, potential bias in the dataset, or the risk of misclassification and its impact on HR decisions.
- The study fails to acknowledge the practical challenges of implementing such a model in real HR environments, where employee decisions are influenced by complex, non-quantifiable factors.
-
Are the references appropriate?
- The study cites many outdated or redundant sources, but lacks references to recent works on transformer-based models and explainable AI in HR analytics.
- It does not reference any real-world industry applications of such models, making it feel disconnected from practical HR concerns.
- No comparison with alternative approaches used in industry—most HR departments rely on interpretable models rather than deep learning, and this aspect is completely ignored.
-
Additional comments on tables and figures:
- ROC and accuracy graphs lack clarity—the results do not provide confidence intervals, making it difficult to assess their reliability.
- Tables are too simplistic—there is no statistical significance testing to validate whether the improvements over other models are meaningful.
- The loss/accuracy curves suggest overfitting, but the authors do not address how they controlled for this issue.
Final Assessment:
While the paper presents a seemingly impressive deep learning model for predicting employee attrition, it suffers from serious methodological flaws, a lack of real-world validation, and questionable improvements over existing approaches. The study does not fill a significant research gap, and its high accuracy claims are likely exaggerated due to overfitting and data augmentation biases. Moreover, the absence of interpretability and practical implications makes it unsuitable for real HR applications.
Key issues that need to be addressed before this work can be considered a meaningful contribution:
- Conduct benchmarking on multiple real-world datasets to verify generalizability.
- Compare the model with state-of-the-art transformers and explainable AI techniques.
- Provide more transparency in hyperparameter tuning and evaluation metrics.
- Incorporate real-world constraints and HR decision-making aspects, not just raw accuracy numbers.
- Address potential biases introduced by GAN-generated synthetic data.
1. Limitations of the Data Sample
Recommendation:
- Conduct experiments using datasets from various industries, such as government institutions, IT companies, and manufacturing enterprises.
- Utilize data from diverse geographical regions to enhance the generalizability of the findings.
2. Insufficient Use of Textual Data
Recommendation:
- Incorporate textual data analysis to capture additional insights.
- Include subjective factors that may influence employee attrition.
3. Further Improvement of the Bi-TCN Model
The study does not analyze which features are the most significant for predicting employee attrition.
Recommendation:
- Perform feature importance analysis using interpretable methods.
- Identify the key factors that have the greatest impact on employee turnover.
4. Lack of Comparison with Alternative Data Balancing Methods
Recommendation:
- Include a comparison with classical data balancing techniques.
- This would ensure that the superior performance of GAN is due to its effectiveness rather than merely the increased dataset size.
Conclusion
- The study represents a significant contribution to the field of employee attrition prediction.
- However, broader validation on various industrial datasets, as well as the incorporation of unstructured information, is required to further strengthen the findings.
Author Response
Please see the attached file.
Author Response File: Author Response.docx
Reviewer 3 Report
Comments and Suggestions for AuthorsThis paper deals with an exciting topic. The article has been read carefully, and some minor issues have been highlighted in order to be considered by the author(s).
(1) The manuscript presents a deep learning framework based on a Bidirectional Temporal Convolutional Network (Bi-TCN) for predicting employee attrition. While the study includes performance comparisons with classical machine learning, deep learning models, and state-of-the-art approaches, it lacks a direct experimental evaluation against well-established baseline models specifically designed for employee attrition prediction. A more comprehensive comparison with existing methods in the literature would strengthen the validity of the proposed approach.
(2) Although the proposed Bi-TCN model demonstrates superior predictive performance, the paper does not provide an in-depth analysis of its computational efficiency. A detailed assessment of time complexity (e.g., training and inference time) and space complexity (e.g., memory usage) is necessary to understand the feasibility of deploying the model in real-world organizational settings. Such an analysis would offer valuable insights into the model’s scalability and practical applicability.
(3) The study employs several components, including the Bi-TCN architecture and a GAN-based data augmentation technique. However, it is unclear how each component contributes to the overall performance of the model. An ablation study isolating the effects of key components (e.g., Bi-TCN without data augmentation, alternative augmentation techniques, different Bi-TCN configurations) would provide a deeper understanding of their individual contributions and the robustness of the proposed method.
(4) It would be beneficial to briefly introduce related research on AI security (adversarial example) from a security perspective, such as the papers available at https://doi.org/10.3390/electronics14030574
Author Response
Please see the attached file.
Author Response File: Author Response.docx
Round 2
Reviewer 2 Report
Comments and Suggestions for AuthorsAccept in present form
Reviewer 3 Report
Comments and Suggestions for AuthorsI recommend the acceptance.
Comments on the Quality of English LanguageN/A