Student retention in higher education remains a significant challenge despite decades of research. This study introduces a novel cluster-label-classify methodology to predict at-risk students and identify common characteristics among those who drop out. Using data from a small private Midwestern university in the United States, we first applied the K-Prototypes algorithm to cluster non-retained students into five groups based on both numeric and categorical variables. We then labeled these clusters and the retained students, creating a multi-class classification problem. Finally, we used a Gradient Boosting Classifier and XGBoost for classification, achieving F1 scores of 0.82 to 0.89 for predicting non-retained students and 0.96 for retained students after addressing class imbalance with SMOTE. This approach allows for customized labeling specific to each institution and enables more targeted interventions for at-risk students. Our methodology combines demographic, academic, and socioeconomic factors to provide a comprehensive view of student retention, potentially offering new insights into this longstanding issue in higher education. The paper also discusses algorithmic bias, examining potential fairness issues in the predictive models and their implications for different student populations. Finally, a discussion on Privacy Preserving Machine Learning (PPML) provides future strategies for testing how these technologies generalize to other institutions while enhancing the privacy of student data.

Institutions of Higher Education (HEI) often possess rich text data in the form of student surveys. However, because these data are expensive to process, many universities have not yet capitalized on this resource. When working with student text data, researchers often desire to first label student responses with common categories of interest, a task in Natural Language Processing known as Multi-label Text Classification (MLTC). BERT and other Large Language Models have produced state of the art results on MLTC tasks; yet because MLTC generally presents challenges of data scarcity and data sparsity, accuracy often remains too low to fully automate the task. Unlike many common MLTC datasets, these student survey data can usually be paired with rich tabular data, both academic and demographic. In this paper, we show that a fusion approach combining tabular data with BERT outputs derived from student responses significantly improves model performance, increasing label ranking average precision from.75 to.84. The paper thus contributes to the open academic discussion of whether fusing tabular demographic data with BERT outputs improves performance, and also offers a practical approach for HEIs to automate survey labeling and thus incorporate more student text data into institutional research.