A Hybrid Unsupervised ML Framework for Predicting Student Admissions in Higher Education: A Case Study

In higher education admissions, particularly in resource-constrained institutions, researchers face challenges in predicting student enrollment likelihood from post-entrance exam visitor data due to imbalanced or unlabeled datasets. This paper proposes a hybrid unsupervised machine learning framework that integrates One-Class Support Vector Machine (OCSVM) for anomaly detection, Gaussian Mixture Model (GMM) for probabilistic clustering, and Nearest Neighbors (NN) for similarity-based scoring. We apply the framework to a real-world dataset of 721 admitted students from D. Y. Patil College of Engineering and Technology (DYPCET), Kolhapur, India. The framework engineers domain-specific features (academic performance, geographic proximity, engagement indicators) to compute a composite likelihood score (0-1 scale). Evaluation yields an 86.1% inlier rate for OCSVM, a silhouette score of up to 0.342 (mean 0.242 ± 0.061 across random initializations) for GMM, and an average NN distance of 0.468. The system enables targeted faculty follow-ups, supporting admission cells in prioritizing high-likelihood candidates for counselor outreach. This deployable pipeline addresses gaps in unsupervised admission prediction for small colleges with limited labeled data.