Predicting Frost Depth of Soils in South Korea Using Machine Learning Techniques
Round 1
Reviewer 1 Report
The authors employed numerous machine learning algorithms to predict the frost depth of soil below the pavement. Their work demonstrate that the GB method can well fitting the physical truth, which can facilitate the prediction of different levels of frost depth and assessment of the sensitivity analysis of pavement predictors. In my opinion, this manuscript is well-written, which employed modern and advanced techniques in this topic and are useful for the researcher who interests in frost depth prediction. I only wonder why the authors choose these 8 algorithms in this study? Only a table is not enough for readers who aren’t familiar with this field therefore, more description should be added.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
General comment: This paper collected a large amount of data and compared the effects of various machine learning methods on frost depth prediction. The result is helpful in the research of frost depth prediction. However, there is not enough discussion on the defects of the data itself and the applicable conditions of the machine learning model. In addition, there are obvious deficiencies in the detailed description of hyperparameter tuning, which need to be supplemented. Details about these problems are listed below:
1. The dataset used in this paper does not contain data that can directly or indirectly reflect soil properties, and the machine learning method adopted in this paper is data-driven in nature. This will lead to great limitations of the model established in this paper, which should be explained.
2. I notice that k value for cross validation were selected differently during performance measurement and hyperparameter tuning. What are the considerations?
3. In table 4, some machine learning methods differ greatly in training sets and test sets, these models may have been fitted. However, it is necessary to combine the training process curve and other information to have a more accurate judgment. It is suggested that the author carry out corresponding examination.
4. The description of hyperparameter tuning in section 3.4, “Several important hyperparameters for each ML algorithm were first selected followed by the selection of a few numbers in a reasonable range for each hyperparameter.”, is quite vague. Since this is an important part in the paper, the “reasonable range” need to be specified.
5. The performance of XGB method in this article is inferior than the GB method, which is not a common case. In view of the above two comments, I believe there may be some defects in the hyperparameter tuning part that are not reflected in the paper.
6. Frost depth prediction is a regression problem, but confusion matrix is used in section 4.3, which is very abrupt. If there is some practical meaning in using confusion matrix for evaluation, it should be further explained in the paper.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 2 Report
The revision by the author provides me with better understandings on the details of the manuscript. Besides, the author made clear response to all of my comments. I think the manuscript can be published. There are still some small problems in this paper, which are listed as follows.
1. Comparing Table 3 and Table 4, the performance of SVM become worse after hyperparameter tuning (the R2 gets lower and RMSE gets higher on the testing set). The author should check for the problems.
2. Still in Table 3 and Table 4, in the testing set, some models have increased R2 score and RMSE value after fine-tuning (the models become worse in terms of RMSE value), while some have decreased R2 score and RMSE value. What specific indicator do the authors use to suggest that the performance of models improves after fine-tuning.
3. As Table 5 changes, the description of Figure 6 needs to be changed accordingly.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf