We present H-NGPCA, a hierarchical clustering algorithm for data streams that integrates an adaptive unit number growth and local dimensionality control. Unlike existing algorithm, H-NGPCA combines the characteristics of centroid-based, model-based and hierarchical clustering. H-NGPCA builds a hierarchical structure of local Principal Component Analysis (PCA) units, where each unit is a hyper-ellipsoid whose shape is updated by a neural network-based online PCA method. The re-positioning of each unit is handled by Neural Gas, a centroid-based clustering algorithm. In the hierarchical tree structure, new units are created in a branch if suggested by a splitting criterion. In addition, each unit determines its own dimensionality based on the data represented by the unit. In extensive benchmarks, H-NGPCA not only surpasses all competing online algorithms with adaptive unit numbers but also achieves competitive performance with state-of-the-art offline methods, reaching an average NMI = 0.87 and CI = 0.26. This demonstrates that H-NGPCA achieves both online adaptability and offline-level accuracy.
Migenda et al. (Mon,) studied this question.