Table 1 Common feature selection algorithm
From: Predictive modelling of thermal conductivity in single-material nanofluids: a novel approach
Feature selection type | Brief description of the feature selection algorithm | Selected features and their importance | Model performance | |
|---|---|---|---|---|
Minimum redundancy maximum relevance (MRMR) | The MRMR (minimum redundancy maximum relevance) algorithm aims to identify an optimal feature subset highly relevant to the response variable and maximally dissimilar | VF = 0.2279, TC = 0.2000 | RMSE (Validation) | 5.65 |
MSE (Validation) | 31.96 | |||
RSQUARED (Validation) | 0.40 | |||
MAE (Validation) | 4.57 | |||
MAE (Test) | 5.34 | |||
MSE (Test) | 39.61 | |||
RMSE (Test) | 6.29 | |||
RSQUARED (Test) | 0.25 | |||
FTest (Importance > 25) | This involves conducting separate Chi-square tests for each predictor variable to determine if there is a significant association between the predictor and the response | VF = 50.7728, DP = 31.8726, NPk = NPa = NPcp = NPmp = NPri = NPek = NPms = NPd = 26.5023 | RMSE (Validation) | 3.50 |
MSE (Validation) | 12.27 | |||
RSQUARED (Validation) | 0.78 | |||
MAE (Validation) | 2.64 | |||
MAE (Test) | 2.33 | |||
MSE (Test) | 9.72 | |||
RMSE (Test) | 3.12 | |||
RSQUARED (Test) | 0.77 | |||
RReliefF (> Abs (0.01)) | The RReliefF algorithm considers the consistency of predictor values among neighbours with the same response values. It penalises predictors exhibiting inconsistent values among neighbouring instances with the same response while rewarding predictors demonstrating differing values among neighbours with different response values | VF = 0.1515, BFv = 0.0142, BFkv = 0.0126, DP = − 0.0092 | RMSE (Validation) | 2.66 |
MSE (Validation) | 7.07 | |||
RSQUARED (Validation) | 0.86 | |||
MAE (Validation) | 1.94 | |||
MAE (Test) | 1.95 | |||
MSE (Test) | 7.39 | |||
RMSE (Test) | 2.72 | |||
RSQUARED (Test) | 0.90 | |||