Transform Residual K-Means Trees for Scalable Clustering

The K-means problem, i.e., to partition a dataset into K clusters is a fundamental problem common to numerous data mining applications. As it is an NP-hard problem, iterative optimizations are typically used such as the K-means algorithm to compute cluster centers. As the K-means algorithm is exponential both in computation and storage in terms of required bits to encode cluster centers, approaches with low computation and storage complexity have been actively studied both for signal compression under real-time constraints and for clustering of large scale high-dimensional datasets. By generating cluster centers via Cartesian products of cluster centers in multiple groups or multiple stages, product and residual K-means trees reduce computation and storage complexity, making it possible to cluster large scale datasets. As residual K-means trees do not require assumed statistical independence among groups that are required by product K-means trees, they generally give better clustering quality. A known limitation of residual K-means trees is that the gain diminishes with added stages. In this paper, by identifying increased intrinsic dimensions of residual vectors with more stages as a limiting factor of performance due to the curse of dimensionality, we propose transform residual K-means trees as a generalization by applying cluster specific transforms to increase the correlations of residual vectors. Our methods substantially reduce the increase of intrinsic dimensions and therefore increase the effectiveness of multiple-stage residual Kmeans trees. Experimental and comparative results using widely used datasets show our methods give the best performance among all scalable methods. Furthermore, our methods enable effective super-clustering learning and provide scalable solutions to duplicate detection and other data mining problems.