Pre-statistics of data for nodes of decision tree

Abstract

Embodiments of the disclosure relate to decision tree generation based on pre-statistics of data for nodes. A plurality of data samples for nodes of a decision tree are obtained, the data samples having respective feature values with respect to a first feature. A target range is determined from a plurality of predefined ranges of values such that the number of feature values falling within the target range exceeds a predetermined threshold number. Then, the remaining feature values other than the feature values falling within the target range are assigned to the corresponding range of values,and the feature values falling into all the ranges of values are counted based on the allocation of the remaining feature values so as to be used for the allocation of the plurality of data samples tochild nodes of the nodes. In this way, the data processing speed and efficiency are significantly improved, and therefore the generation speed and efficiency of the decision tree are increased.

Description

Background technique [0001] Decision trees are a technique widely used in machine learning models or processes. Using this technique, non-linear dependencies between data can be modeled and results can be interpreted without additional feature preprocessing such as normalization. When combined with different loss functions, decision trees can be used across various domains such as classification, recursion, and ranking. Moreover, when combined with different integration techniques such as bagging and boosting, a variety of decision tree algorithms can be derived, including random forests, gradient boosting decision trees (GBDT), and so on. As examples, decision trees have been widely used in network (eg, Web) applications such as document ranking in Web search, click prediction for advertising targets, etc. by combining with different loss functions and different integration techniques. [0002] In the decision tree algorithm, the fitting of a single tree is realized by recu...

AI Technical Summary

Problems solved by technology

The generation of traditional decision trees requires a considerable number of visits to the training data. Such frequent visits will inevitably introduce a large processing delay and greatly reduce the processing efficiency.

Images