ScholarOne - Unlocking Cognitive Patterns: A Comparative Exploration of
Linear and Deep Dimensionality Reduction Approaches in clusterizing
Student Cognitive Profiles.
Abstract
Cognitive profiling plays a crucial role in understanding learning
dynamics, it contributes significantly to the development of students’
metacognitive skills and awareness of the learning process, thereby
facilitating the adoption of tailored learning experiences. Clustering,
proves effective in cognitive profiling. However, the challenge of the
“curse of dimensionality” introduces complexities that can impact the
accuracy of cluster subject attribution. This paper investigates the
evaluation of various cluster internal validation metrics and cluster
stability using a dataset of 1626 participants comprising 54 items
across six cognitive domains from the digital assessment tool, PROFFILO.
We employ three clustering procedures—K-means, Gaussian Mixture
Models, and Fuzzy-C Means—on raw data and apply linear (Principal
Component Analysis) or non-linear (Variational Autoencoders), or a
combination of PCA and VAE dimensionality reduction techniques. Results
indicate that, for high-dimensional cognitive domains, a combination of
PCA and VAE yields superior clustering quality. Conversely, in less
high-dimensional domains, the VAE outperforms the PCA approach. In
summary, the application of dimensionality reduction techniques
demonstrates promising outcomes in student cognitive profiling,
especially for data characterized by high dimensionality and
heterogeneity. These findings have practical implications for advancing
personalized learning experiences and enhancing our understanding of the
intricate relationships within students’ cognitive domains.