Performance Assessment of Machine Learning Models Through Stratified and K Fold Cross Validation

Abstract

Cross validation is a statistical method that partitions a dataset by reserving a portion for model testing (validation set), while the remaining data is utilized for training. Among its variants is the leave one out cross validation technique. Several factors motivate the adoption of K Fold cross validation. Although large datasets can reveal hidden patterns and provide extensive insights, smaller datasets pose challenges when split into training and testing subsets at ratios such as 80:20, as this division may not yield sufficient evaluation power. In contrast, K Fold cross validation ensures that all portions of the dataset are eventually used for testing, thereby maximizing data utility. Variants include K Fold, Stratified K Fold, Time Series cross validation, and leave one out cross validation. In the proposed work, K Fold cross validation was employed to partition the dataset into training and testing subsets, followed by model evaluation to determine accuracy. The objective was to assess whether the model exhibited overfitting or underfitting, thereby establishing its overall quality. By applying K Fold cross validation, the model was tested across k different subsets, which enhanced generalizability and facilitated the selection of the model with the best performance.