Evaluating Recommendation System Bias on a Public Dataset based on Demographic Information

Abstract

Technological developments play a pivotal role in shaping the future of a complex and changing world. Student dropouts significantly impede a country’s economic growth and society, and students miss out on employment opportunities. Advances in artificial intelligence have improved the prediction of student success and dropouts. The training data often perpetuates algorithm bias. The study used a public dataset containing demographic, socioeconomic (sensitive data) and academic data to predict success or dropout risks. The study investigates the impact of sensitive data on the performance of algorithms. Results show that algorithms that use sensitive data perform slightly better than algorithms that do not. The light gradient boosting machine and the random forest were the most effective algorithms when using the entire dataset and a dataset without sensitive data, respectively. Our study contributes to the growing literature on challenges with bias in recommender systems. Future studies could use actual institutional data to build algorithms and use additional sensitive information for improved results.