Comparison of Clustering Approaches for Anomaly Detection on Internet of Things Data

Clustering approaches are useful in gathering similar data together. Here, we look at K-Means, DBSCAN, H-DBSCAN, and Expectation Maximization using Gaussian Mixture Modeling to group IoT environmental data into anomalous and non-anomalous clusters. We perform the analysis of these algorithms using confusion matrices based on defining anomalies heuristically, and study the effectiveness of cluster grouping visually through graphical analysis. Results show Expectation Maximization outperforms all other algorithms. Additionally, further analysis shows the dynamic algorithms used perform better than reported in earlier research.