Increasing Accuracy in Predicting Student Test Scores with Regression Using Domain Reduction Technique of Principal Component Analysis

It is difficult to predict student test scores and a number of algorithms are often used for such prediction indicating the importance of such research. To improve the accuracy of predicting student test scores, this research combines Principal Component Analysis (PCA) and Regression. PCA is used to reduce the size of the domain before being used in Regression. For this experimental study, the test scores for both Mathematics and Language were extracted from a public dataset. Test scores were then forecasted using Polynomial Regression. The calculated Mean Absolute Error was used as a criterion of measure for accuracy. Domain values were then subjected to PCA using the same splitting of training and test data, and Regression was also used to train the new set of domain values. Utilizing PCA, this process was repeated for all number of attributes of the domain. Training and test data for the datasets were randomly selected. Each number of domain values was tested in 50 trials. An average mean absolute error was calculated. A statistical T-Test was performed to determines if difference between the Mean Absolute Error recorded were statistically significant. The threshold used to determine if the results were not a statistical coincidence was a p-value of 0.05. Results of using PCA in conjunction with Regression have decreased average mean error substantially. For both >0.00001and Languages test scores using PCA can decrease mean absolute errors by more than 50%. But not all domain sizes improved accuracy. Some produce worse results than not using PCA. This is similar to previous published research. This research demonstrates that PCA can significantly improve student test score prediction when using Regression. Results were comparable to previous research that showed PCA can increase accuracy when using Neural Networks for this problem. PCA is a promising technique to increase the accuracy of such predictive models.