This study explores data mining techniques for predicting student dropout in higher education. The research compares different methodological approaches, including alternative algorithms and variations in model specifications. Additionally, we examine the impact of employing either a single model for all university programs or separate models per program. The performance of models with students grouped according to their position on the program study plan was also tested. The training datasets were explored with varying time series lengths (2, 4, 6, and 8 years) and the experiments use academic data from the University of Porto, spanning the academic years from 2012 to 2022. The algorithm that yielded the best results was XGBoost. The best predictions were obtained with models trained with two years of data, both with separate models for each program and with a single model. The findings highlight the potential of data mining approaches in predicting student dropout, offering valuable insights for higher education institutions aiming to improve student retention and success.