Hybrid Embedding and Clustering techniques for Enhanced Stress Detection

Stress detection from physiological signals has gained significant attention due to its implications for mental health and human well-being. Traditional machine learning approaches rely heavily on supervised methods, which require extensive labeled data, limiting their applicability in real-world scenarios. In this study, we systematically investigate the impact of Dimensionality Reduction (DR) techniques on unsupervised stress detection. We apply four DR methods, including PCA, UMAP, t-SNE, and t-SNE-PSO, to extract meaningful representations of benchmark stress datasets. These embeddings serve as inputs for four clustering algorithms, k-Means, k-medoids, Agglomerative Clustering, and Spectral Clustering, to evaluate their effectiveness in stress-state separation. Additionally, we propose hybrid embedding strategies that combine components from different DR techniques, further enhancing clustering performance. The experimental results demonstrate the effectiveness of the DR techniques and their hybrid embeddings in achieving the best clustering performance and visualization in terms of objective and subjective evaluations. This study highlights the importance of selecting appropriate embeddings for clustering-based stress detection and underscores the potential of hybrid representations in uncovering meaningful stress patterns.