Using Factor Analysis to Increase Accuracy of Neural Networks for Predicting Student Test Scores

This work is a full research-to-practice paper that describes a predictive method to improve the prediction of student test scores. Predicting student test scores is difficult. However, doing so can improve education greatly by improving advising, scheduling, tutoring assignment and other educational processes. This research extends previous research by using a domain space reduction technique to improve accuracy. Factor Analysis is used to reduce the number of domain attributes for improving the accuracy of a Neural Network to predict student test scores. In this research datasets for Mathematics and Language of high school student test scores were used. Test scores were predicted using a Neural Network computing the Mean Absolute Error as a measurement of accuracy. The datasets have 30 domain attributes each. Factor Analysis was used to reduce the domain size from between 1 to 29, each time using it to train the Neural Network. Because the Mean Absolute Error may vary depending upon which records in the dataset are used for training versus testing, 50 trials of each dataset size were executed producing an Average Mean Absolute Error for each domain size. A statistical test was used to show statistical significance between the Neural Network without Factor Analysis and the Neural Network with varying domain sizes using Factor Analysis. Results were very promising and correspond to previous research that used Principal Component Analysis. Numerous domain sizes had significantly better Average Mean Absolute Errors than the accuracy of the Neural Network without Factor Analysis. This research shows that reducing the domain size using Factor Analysis can greatly improve the accuracy of Neural Networks when predicting student test scores. The best improvements occurred when domain sizes were very small ranging from 2 to 6. Domain reduction techniques, such as Factor Analysis, have been shown to improve predictive models for student test score prediction. Future research should consider other domain reduction techniques, like T-SNE, in combination with other predictive models, like Support Vector Machines.