Fig 1.
An example of the relationship between a factor and its observed features.
Fig 2.
An example of using a component to represent its corresponding features.
Fig 3.
An example of the process of extracting signals from the cocktail party problem with two speaking people (source signals) and two microphones (mixture signals).
Fig 4.
An example of MLP with three input neurons, two hidden neurons, and one output neuron.
Fig 5.
ε-SVM regression with the ε-insensitive hinge loss, meaning there is no penalty to errors within the ε margin.
Fig 6.
An example of the RF model.
Fig 7.
An example of the XGBoost model.
Table 1.
Parameter settings for the prediction models, where #neurons is the number of neurons, #iterations is the maximum number of iterations, regularisation is the regularisation parameter, σ2 is the variance within the RBF kernel, #trees is the number of trees, and depth is the maximum depth of the tree.
Fig 8.
A flowchart of different feature extraction methods used for body fat prediction based on K-fold cross validation with N repeated experiments.
Table 2.
Statistical properties of Case 1’s body fat dataset.
Fig 9.
Explained variance ratio for the StatLib dataset.
Table 3.
Experimental results based on the StatLib dataset (best results are highlighted in bold).
Table 4.
Wilcoxon rank-sum tests for the MLP, SVM, RF, XGBoost, and the use of feature extraction, based on the StatLib dataset in terms of RMSE (p-values less than 0.05 are highlighted in bold).
Table 5.
Experimental results for the MLP, SVM, RF, and XGBoost, based on the StatLib dataset, with FA feature extraction (best results are highlighted in bold; # means the number of features).
Table 6.
Experimental results for the MLP, SVM, RF, and XGBoost, based on the StatLib dataset, with PCA feature extraction (best results are highlighted in bold; # means the number of features).
Table 7.
Experimental results for the MLP, SVM, RF, and XGBoost, based on the StatLib dataset, with ICA feature extraction (best results are highlighted in bold; # means the number of features).
Fig 10.
Comparison results in terms of computation time based on FA, PCA and ICA feature extraction for the StatLib dataset.
Table 8.
Statistical properties of Case 2’s body fat dataset.
More details can be found at https://www.cdc.gov/nchs/nhanes/index.htm.
Fig 11.
Explained variance ratio for the NHANES dataset.
Table 9.
Experimental results based on the NHANES dataset (best results are highlighted in bold).
Table 10.
Wilcoxon rank-sum tests for the MLP, SVM, RF, XGBoost, and the use of feature extraction, based on the NHANES dataset in terms of RMSE (p-values less than 0.05 are highlighted in bold).
Table 11.
Experimental results for the MLP, SVM, RF, and XGBoost, based on the NHANES dataset, with FA feature extraction (best results are highlighted in bold; # means the number of features).
Table 12.
Experimental results for the MLP, SVM, RF, and XGBoost, based on the NHANES dataset, with PCA feature extraction (best results are highlighted in bold; # means the number of features).
Table 13.
Experimental results for the MLP, SVM, RF, and XGBoost, based on the NHANES dataset, with ICA feature extraction (best results are highlighted in bold; # means the number of features).
Fig 12.
Comparison results in terms of computation time based on FA, PCA, and ICA feature extraction for the NHANES dataset.