Increasing Accuracy in Predicting Student Test Scores with Neural Networks using Domain Reduction Technique of Principal Component Analysis

Abstract

This research uses Principal Component Analysis (PCA) in conjunction with a Neural Network to increase the accuracy of predicting student test scores. Much research has been conducted attempting to predict student test scores using a standard, well-known dataset. The dataset includes student demographic and educational data and test scores for Mathematics and Language. Multiple predictive algorithms have been used with a Neural Network being the most common.In this research PCA was used to reduce the domain space size using varying sizes. This began with just 1 attribute and increased to the full size of the original set’s domain values. The reduced domain values and the original domain values were independently used to train a Neural Network and the Mean Absolute errors were compared. Because results may vary depending upon which records in the dataset are training versus testing, 50 trials were conducted for each reduction size. Results were average and statistical tests were applied. Results show that using PCA prior to training the Neural Network can decrease the mean absolute error by up to 15%.