UNSUPERVISED EXPLORATION OF HIERARCHICAL STRUCTURE IN HUMAN ACTIVITY RECOGNITION DATA

Abstract

Human Activity Recognition (HAR) datasets contain complex patterns that supervised models exploit with labeled training, but it remains unclear what latent structure exists in the data itself. This paper presents an unsupervised exploratory analysis of a smartphone sensor HAR dataset to uncover inherent activity groupings without using activity labels. We apply a range of clustering algorithms (k-means, Gaussian mixture, hierarchical agglomerative, density-based HDBSCAN, spectral clustering) and dimensionality reduction methods (Principal Component Analysis – PCA, t-distributed Stochastic Neighbor Embedding – t-SNE, Uniform Manifold Approximation and Projection – UMAP, and a feed-forward autoencoder) to identify natural clusters of sensor feature vectors. Quantitatively, we evaluate clustering quality using internal metrics (silhouette coefficient) and external metrics against true labels (Adjusted Rand Index – ARI, and Normalized Mutual Information – NMI). The results reveal a dominant two-cluster division separating static postures from dynamic movements, with finer sub-clusters roughly corresponding to the six known activities when clustering is applied hierarchically. UMAP non-linear embedding dramatically improved cluster separability and alignment with classes, outperforming PCA. Analyzing feature importance in each cluster showed that features related to body orientation and acceleration dynamics differentiate activities. These findings demonstrate that unsupervised learning can automatically discover meaningful activity groupings (e.g. distinguishing stationary vs. moving behaviors) and key distinguishing sensor features, without any labels. The study provides insights into intrinsic HAR data structure, which can inform feature design and hierarchical modeling in future activity recognition systems.