Detection of Fault Events in Software Tools Integrated with Human–Computer Interface Using Machine Learning

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

My review about the paper is brief, but I think it is enough which I am going to comment:

1）Please use an algorithm form to outline the detailed learning steps instead of a brief description using texts.

2）The paper could be improved if the authors can provide the time complexity analysis of the proposed algorithm. If it is difficulty, I would like to see the comparison results in terms of the running time.

3）As a peer-reviewed paper, some very related references are missing, mainly including some previously published papers on Defect Detection:  Cost-Sensitive Residual Convolutional Neural Network for PCB Cosmetic Defect Detection

4）It is more convinced if the authors can compare the proposed algorithm with some other state-of-the-art algorithms.

5）Please format the paper using the required article class. The typography of the current version is very poor.

6）My last concern is the language usage. The authors should at first proofread the whole paper and then try to seek some help on improving the English quality of their manuscript.

Author Response

Reviewer 1

Comment 1: “Please use an algorithm form to outline the detailed learning steps instead of a brief description using texts.”
Response: Addressed. In Section 3.7 (Algorithm Representation), we added Algorithm 1 to outline the full training and evaluation steps of the RBM–LR model in a clear algorithmic format.

Comment 2: “The paper could be improved if the authors can provide the time complexity analysis of the proposed algorithm. If it is difficult, I would like to see the comparison results in terms of the running time.”
Response: Addressed. We included a Time Complexity Analysis in Section 3.8, with complexity expressed as O(n × h × f × epochs). In addition, runtime comparisons were included in Section 4.5 (Runtime Analysis).

Comment 3: “Some very related references are missing, mainly including previously published papers on defect detection.”
Response: Addressed. We added several recent references (2024–2025), including studies in HCI, ML, and defect prediction (Du et al., Yuan et al., Dayanand et al., Sadeghi Milani et al., Sun et al.) in the References section.

Comment 4: “It is more convincing if the authors can compare the proposed algorithm with some other state-of-the-art algorithms.”
Response: Addressed. In Section 4.4 (Experimental Evaluation) and 4.7 (Summary of Findings), we compared RBM–LR against Random Forest, XGBoost, and SVM, and demonstrated superior results.

Comment 5: “Please format the paper using the required article class. The typography of the current version is very poor.”
Response: Addressed.

Comment 6: “The authors should proofread the paper and improve the English language.”
Response: Addressed. The entire paper underwent careful language editing to improve clarity, grammar, and readability. The final version is significantly more polished compared to the original

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This paper is devoted to the development and evaluation of the RBM-LR approach for classifying user errors in software interaction. The authors propose a combination of feature extraction methods using Restricted Boltzmann Machine (RBM) and classification using logistic regression (LR). The effectiveness of this approach is confirmed on several datasets, where RBM-LR demonstrates higher classification accuracy compared to existing methods.

Сomments on the article:

1. “The outline of the paper is presented below: Section 2 provides the related works. Section 3 provides the details of the proposed model. Section 5 evaluates the entire work and section 6 concludes the work.” Please indicate what is contained in section 4
2. The co-authors use RBM for feature extraction, but there is no comparative analysis with other dimensionality reduction or feature transformation methods (e.g., PCA, Autoencoder, UMAP). Since RBM requires careful tuning of hyperparameters and does not always perform better than other current methods.
3. Logistic regression as a simple basic classifier is often used as a starting point. But if the article positions the approach as “new” or “optimized”. Therefore, it is necessary to explain why more advanced algorithms (Random Forest, XGBoost, SVM, , Neural Networks) were not used
4. The text only mentions that AUC is used as a target, but does not explain how it is optimized. Therefore, it is worth explaining whether AUC is included in the loss function or is only used as a metric after training.
5. “Figure 5 shows the confusion matrix that can be used to compute the AUC.” Figure 5 is missing
6. Unclear motivation for choosing RBM-LR. The authors did not justify why the combination of RBM and logistic regression is better for the SFP problem compared to other modern deep or ensemble models (e.g., XGBoost, LSTM, Autoencoder).
7. There is no analysis of the model training time and computational costs of RBM-LR compared to the baseline LR or other models.
8. There is no verification of the statistical significance of the results ( t-test or analysis of variance ) of RBM-LR compared to the baseline LR. This reduces the confidence in the interpretations.
9. Conclusion section should be extended using: 1) numerical results obtained in the paper; 2) limitations of the proposed approach; 3) prospects for the future research.
10. It is advisable to provide a link to the code or pseudocode of the model on open resources, such as GitHub, for reproducibility of the results of RBM-LR implementation.

Comments on the Quality of English Language

This paper is devoted to the development and evaluation of the RBM-LR approach for classifying user errors in software interaction. The authors propose a combination of feature extraction methods using Restricted Boltzmann Machine (RBM) and classification using logistic regression (LR). The effectiveness of this approach is confirmed on several datasets, where RBM-LR demonstrates higher classification accuracy compared to existing methods.

Сomments on the article:

1. “The outline of the paper is presented below: Section 2 provides the related works. Section 3 provides the details of the proposed model. Section 5 evaluates the entire work and section 6 concludes the work.” Please indicate what is contained in section 4
2. The co-authors use RBM for feature extraction, but there is no comparative analysis with other dimensionality reduction or feature transformation methods (e.g., PCA, Autoencoder, UMAP). Since RBM requires careful tuning of hyperparameters and does not always perform better than other current methods.
3. Logistic regression as a simple basic classifier is often used as a starting point. But if the article positions the approach as “new” or “optimized”. Therefore, it is necessary to explain why more advanced algorithms (Random Forest, XGBoost, SVM, , Neural Networks) were not used
4. The text only mentions that AUC is used as a target, but does not explain how it is optimized. Therefore, it is worth explaining whether AUC is included in the loss function or is only used as a metric after training.
5. “Figure 5 shows the confusion matrix that can be used to compute the AUC.” Figure 5 is missing
6. Unclear motivation for choosing RBM-LR. The authors did not justify why the combination of RBM and logistic regression is better for the SFP problem compared to other modern deep or ensemble models (e.g., XGBoost, LSTM, Autoencoder).
7. There is no analysis of the model training time and computational costs of RBM-LR compared to the baseline LR or other models.
8. There is no verification of the statistical significance of the results ( t-test or analysis of variance ) of RBM-LR compared to the baseline LR. This reduces the confidence in the interpretations.
9. Conclusion section should be extended using: 1) numerical results obtained in the paper; 2) limitations of the proposed approach; 3) prospects for the future research.
10. It is advisable to provide a link to the code or pseudocode of the model on open resources, such as GitHub, for reproducibility of the results of RBM-LR implementation.

Author Response

Reviewer 2

Comment 1: “Please indicate what is contained in Section 4.”
Response: Addressed. We added Section 4 (Results and Discussions) with detailed experiments, evaluations, and comparisons.

Comment 2: “The co-authors use RBM for feature extraction, but there is no comparative analysis with other dimensionality reduction or feature transformation methods (e.g., PCA, Autoencoder, UMAP).”
Response: Addressed. In Related Works, we provided a comparative discussion explaining why RBM was selected over PCA, Autoencoders, and UMAP.

Comment 3: “Logistic regression as a basic classifier is often used as a starting point. Please explain why more advanced algorithms (Random Forest, XGBoost, SVM, Neural Networks) were not used.”
Response: Addressed. In Introduction, we justified the use of Logistic Regression (LR) due to its interpretability, computational efficiency, and suitability for industrial applications compared to more complex models.

Comment 4: “The text only mentions that AUC is used as a target but does not explain how it is optimized.”
Response: Addressed. In 3.2 (Objective Function) and 4.2 (Performance Measures), we clarified that AUC was used only as an evaluation metric, not as part of the loss function.

Comment 5: “Figure 5 (confusion matrix) is missing.”
Response: Addressed. In Section 4.2, we included and explained the confusion matrix used to compute Precision, Recall, F1, and AUC.

Comment 6: “Unclear motivation for choosing RBM–LR over other deep/ensemble models.”
Response: Addressed. In Introduction and Related Works, we emphasized that RBM–LR strikes a balance between interpretability and performance, making it suitable for defect prediction tasks.

Comment 7: “There is no analysis of the model training time and computational costs.”
Response: Addressed. In 3.8 Time Complexity Analysis and 4.5 Runtime Analysis, we presented a detailed runtime evaluation and complexity formula.

Comment 8: “There is no verification of statistical significance (t-test or ANOVA).”
Response: Addressed. In Section 4.6, we included paired t-tests across datasets, showing results are statistically significant (p < 0.05).

Comment 9: “The conclusion should be extended with numerical results, limitations, and future prospects.”
Response: Addressed. The Conclusion now includes:

• Numerical improvements (7% gain in AUC, 6% in accuracy, etc.),
• Limitations (sensitivity to hyperparameters, runtime overhead),
• Future directions (integration with LSTM/Transformers, industrial datasets, explainability methods).

Comment 10: “It is advisable to provide pseudocode or code link for reproducibility.”
Response: Addressed. We included Algorithm 1 pseudocode and stated that the code will be made available on GitHub.

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The co-authors took into account all the previous comments, so I believe it can be accepted for publication as is.

Comments on the Quality of English Language

The co-authors took into account all the previous comments, so I believe it can be accepted for publication as is.

Author Response

Comment 1. Please include some missing related work citations on PCB cosmetic defect detection

Response 1. The PCB cosmetic defect detection methods have been added to the related work section, and the corresponding reference has been added.

Comment 2: Give the manuscript a careful language edit.

Response 2: The manuscript was edited carefully, and grammatical checks were carried out.