Educational data mining is a growing field in education, but there is no consensus on the interpretability of data mining algorithms. We propose a student feature selection strategy based on the random forest algorithm that scales the gap between two features and adopts a geometric mean method to balance the contribution of performance importance and robustness, which is useful for educators and academics Researchers are vital. We collected a dataset of academic performance and characteristics of students from Guangdong Vocational College of Finance and Trade for experiments. Experiments show that our feature selection strategy is highly effective and produces relatively great results in both random forest and linear regression algorithms. Additionally, our student feature selection strategy provides valuable insights into the field of educational data mining by offering a clearer understanding of the factors influencing students ' academic performance.