Integration of Stratified KFold Cross Validation to Enhance Prediction Accuracy: A Comparison Study

Several researchers explore the impact of using KFold cross-validation in enhancing the model’s prediction performance. Improvement in the accuracy was shown in the available presented researches’ findings. The previous study focuses on feature selection methods and comparison of algorithms to declare the best-performing predictive model in forecasting students’ success to graduate from a real-world dataset in a 70/30 split. To further develop a better predictive model, this study integrated stratified KFold cross-validation to the same binary encoded dataset but an 80/20 split. The result shows that in NFS, 4/7 classifiers increased accuracy and L2 only 3/7. Meanwhile, in L1 feature selection, all seven (7/7) classifiers were seen to improve accuracy and decision Tree with a 14% increase compared to those not using stratified KFold cross-validation. On the other hand, only 3/7 in Genetic Algorithm (GA) with 5/7 improvement in AUC. However, there was a 28% decrease in accuracy using Naïve Bayes. The findings revealed that with the L1 selection method, seven classifiers performed better integrated with cross-validation than none. Moreover, GA and Random Forest earned the highest accuracy (0.79) and AUC (0.75), 4% higher than GA and Logistic without cross-validation.