UMAP-Based Dimensionality Reduction for Educational Data Analysis: Demonstrating Visualization-Informed Clustering Through Student Classification

• In Event: The Nuance of Patterns: Methods for Uncovering Intricate Quantitative Relationships

Abstract

Educational researchers increasingly rely on clustering analysis to understand student diversity, yet traditional approaches select analytical methods based on algorithmic convenience rather than examining natural data characteristics. This study introduces visualization-informed method selection, using Uniform Manifold Approximation and Projection (UMAP) to reveal data structure before choosing clustering techniques. Analyzing behavioral data from 164,313 California community college students, we demonstrate how this approach uncovers authentic diversity patterns potentially obscured by conventional methods. Results reveal a distinctive core-periphery structure with 61% of students following mainstream pathways alongside multiple smaller specialized groups—challenging assumptions about balanced student diversity. This methodological advancement enables evidence-driven rather than assumption-driven analytical decision-making, offering educational researchers enhanced capabilities for understanding complex student populations while maintaining interpretive accessibility.

Authors

• Yiran Chen, University of Michigan

• Ying Sun, University of Michigan

• Peter Riley Bahr, University of Michigan