Isotropic Dynamic Hierarchical Clustering (1605.07030v1)

Abstract: We face a need of discovering a pattern in locations of a great number of points in a high-dimensional space. Goal is to group the close points together. We are interested in a hierarchical structure, like a B-tree. B-Trees are hierarchical, balanced, and they can be constructed dynamically. B-Tree approach allows to determine the structure without any supervised learning or a priori knowlwdge. The space is Euclidean and isotropic. Unfortunately, there are no B-Tree implementations processing indices in a symmetrical and isotropical way. Some implementations are based on constructing compound asymmetrical indices from point coordinates; and the others split the nodes along the coordinate hyper-planes. We need to process tens of millions of points in a thousand-dimensional space. The application has to be scalable. Ideally, a cluster should be an ellipsoid, but it would require to store O(n2) ellipse axes. So, we are using multi-dimensional balls defined by the centers and radii. Calculation of statistical values like the mean and the average deviation, can be done in an incremental way. While adding a point to a tree, the statistical values for nodes recalculated in O(1) time. We support both, brute force O(2n) and greedy O(n2) split algorithms. Statistical and aggregated node information also allows to manipulate (to search, to delete) aggregated sets of closely located points. Hierarchical information retrieval. When searching, the user is provided with the highest appropriate nodes in the tree hierarchy, with the most important clusters emerging in the hierarchy automatically. Then, if interested, the user may navigate down the tree to more specific points. The system is implemented as a library of Java classes representing Points, Sets of points with aggregated statistical information, B-tree, and Nodes with a support of serialization and storage in a MySQL database.