Parallel generation method and device of decision tree on the basis of layered strategy

Abstract

The embodiment of the invention provides a parallel generation method and device of a decision tree on the basis of a layered strategy. The decision tree is constructed according to layers downstream from the top; when the splitting attribute and the optimal split point of each node of an nth layer are calculated, a training data set only needs to be scanned for one time, each piece of training data in the training data set is classified into each node of the nth layer according to the filter condition of each node of the nth layer; then, the splitting attribute and the optimal split point of each node are independently calculated to generate the node information of each node; and according to the node information of each node of each layer of the decision tree, constructing the decision tree. In the above method, when the splitting attribute and the optimal split point of each node of each layer are calculated, the number of training data of each category in each splitting interval of each attribute in each node is subjected to parallel statistics according to the filtering condition of each node, so that the splitting attribute and the optimal split point of each node of the same layer can be subjected to parallel computation, and the performance of the decision tree is improved.

Description

technical field [0001] Embodiments of the present invention relate to data communication technologies, and in particular to a method and device for parallel generation of decision trees based on hierarchical strategies. Background technique [0002] Decision Tree (decisionTree) is a main method for classification and prediction, and is used to represent the tree of the corresponding relationship between decision-making and corresponding decision-making results. Each non-leaf node in the tree represents a decision whose value leads to a different decision result (leaf node) or affects subsequent decision choices. The decision tree corresponds to a rule from the root node to the leaf node, and the whole tree corresponds to a set of expression rules. [0003] The generation process of the existing decision tree adopts a recursive method, and the split attribute of each node of the decision tree is calculated sequentially from top to bottom, and the calculation of the split att...

AI Technical Summary

Problems solved by technology

[0004] Since each map in the MapReduce parallelization algorithm is aimed at the attribute level, and the intermediate results of each map need to be written to the disk, for high-dimensional and massive training data, when the number of nodes in the decision tree increases, The performance of the decision tree parallelization algorithm will be severely degraded, or even unbearable

Images