Impact of Data Splitting Techniques on the Performance of a Convolutional Neural Network Facial Recognition Model

The performance of facial recognition models is influenced by the choice of data partitioning strategy. Appropriate data splitting techniques enable more reliable estimation of model generalization and help assess overfitting. In this study, we examine the performance of a convolutional neural network (CNN)-based facial recognition model under four commonly used data splitting approaches: random splitting, stratified splitting, bootstrap validation, and k-fold cross-validation. Experiments are conducted on two benchmark data sets, Labeled Faces in the Wild and the Olivetti Research Laboratory. Model performance is evaluated using accuracy as well as empirical estimates of model bias and variance. The results indicate that k-fold cross-validation provides a more stable performance estimate under the experimental conditions considered, suggesting its suitability as a validation procedure for CNN-based facial recognition models.