On the performance of machine learning models proposed for electrical conductivity classification of carbon nanotube-doped polymer nanocomposite films

This study presents a comparative performance evaluation of four machine learning classifiers, which are artificial neural networks, support vector machines, bagged decision trees and logistic regression, aiming to predict the conductivity regime (insulator or semiconductor) of carbon nanotube (CNT)-doped polystyrene (PS) latex films using a highly constrained dataset. To ensure scientifically valid and reproducible estimates, a leak-free repeated nested 5-fold cross-validation framework was employed together with variance inflation factor analysis to rule out multicollinearity. Initial standard metrics suggested that the high-capacity bagged tree ensemble possessed near-perfect predictive power. However, subsequent threshold-independent global precision–recall pooling and rigorous Y-scrambling permutation tests definitively exposed this as an artifact of fold-specific memorization, with the ensemble demonstrating a critical failure by perfectly memorizing randomized noise. Ultimately, the logistic regression model was found to be the superior and most robust framework. By enforcing a strict linear decision boundary, the logistic model successfully resisted overfitting, demonstrating that its constrained mathematical structure naturally aligns with the abrupt, macroscopic phase transition characteristic of electrical percolation. Research findings reveal that for modest measurement datasets governed by threshold-like physical behavior, simpler linear models offer a superior bias-variance tradeoff, providing a reliable computational tool to significantly reduce trial and error experiments in nanocomposite design.