A Hierarchal Tensor Based Approach with Boosted Regression Trees

Abstract: Within the split-and-merge paradigm for the compressed, continuously updating and data querying of multidimensional spatial data, the original multidimensional spatial tensor data are divided into small blocks according to their spatial data references. Each block is represented in regression tree structure and at each level, all leaf nodes of regression tree are boosted then compressed hierarchically. Then the blocked hierarchical tensor representation combined into a single hierarchical tree as the representation of original data. With a buffered regression tree data structure and boosting, the corresponding optimized operation algorithms, the original multidimensional spatial field data can be continuously compressed, appended, and queried. The new approach with above used regression trees instead of buffered trees and using boosting algorithms to improve the fit ratio, so that the quality of the original data with much low storage costs and faster computational performance.

Keywords: Geo Spatial Data, Tensors, Regression Trees, Gradient Boosting