HAR Unsupervised Clustering
An unsupervised study of natural movement patterns in the UCI Human Activity Recognition dataset.
RepositoryQuestion and data
This project studies whether smartphone sensor features can reveal meaningful movement groups without using activity labels during training. PCA is used for visualization, with the first two components explaining approximately 56.9% of total variance.
The UCI Human Activity Recognition dataset contains features derived from smartphone motion sensors across several human activities.
Method
The feature space is cleaned and standardized before comparing K-Means and Agglomerative Clustering at K = 2, K = 3, and K = 6. Internal evaluation uses silhouette, Calinski-Harabasz, and Davies-Bouldin scores; Adjusted Rand Index and Normalized Mutual Information are used after clustering for label comparison.
Activity labels are not used during clustering. External metrics are applied afterward only to compare the discovered groups with the known activity labels.
Findings
K = 2 produced the strongest mathematical separation. K = 3 gave a more interpretable split among walking-related activities, while K = 6 did not naturally yield six clean clusters. The clearest structure was static or low-motion activity versus dynamic walking-related activity.
The result suggests that the sensor features naturally separate broad movement categories more readily than six fully distinct human-defined classes.