Enhancing Clustering Performance through Benchmarking of Dimensionality Reduction Techniques on Educational Data

This study evaluates the effectiveness of dimensionality reduction techniques in enhancing clustering performance using a tracer study dataset of 500 alumni from UMNU Kebumen, containing 58 variables. The objective was to identify the optimal combination of dimensionality reduction and clustering methods for uncovering patterns in alumni profiles, job search strategies, and employment outcomes. Principal Component Analysis (PCA), Non- Negative Matrix Factorization (NMF), t-Distributed Stochastic Neighbor Embedding (t-SNE), and Uniform Manifold Approximation and Projection (UMAP) were applied, followed by clustering using K-Means, DBSCAN, and Hierarchical Clustering. The findings revealed that NMF achieved the highest clustering quality, particularly with K- Means and Hierarchical Clustering, outperforming PCA. NMF also demonstrated superior compactness with a Calinski-Harabasz Index of 287.96, compared to 125.88 for PCA. While t-SNE and UMAP delivered competitive results, their computational times of 245.8 and 76.5 seconds, respectively, made them less practical for large datasets. The novelty of this study lies in its comprehensive evaluation of dimensionality reduction techniques and the integration of diverse clustering algorithms to assess their interplay. The results provide actionable insights, recommending NMF for accuracy-critical tasks and PCA for time-sensitive applications. Given the increasing volume of high-dimensional educational data, this study highlights the critical need for efficient clustering strategies to extract meaningful insights, ultimately supporting data-driven decision-making in education and workforce planning. Addressing these challenges is essential to optimizing institutional strategies, improving student employability, and enhancing workforce alignment with industry demands.