Quantization method and device of feature distinguishing capability, equipment and medium

Abstract

The invention discloses a feature distinguishing capability quantification method, which comprises the following steps of: performing abnormal value detection on a data set on the basis of a plurality of decision trees which are randomly constructed in advance so as to determine a to-be-explained sample in a plurality of data samples; determining a plurality of path nodes corresponding to the target to-be-explained sample in the target to-be-explained decision tree; acquiring node features and node thresholds at the plurality of path nodes, and constructing a node tree according to the node features, the node thresholds and feature values of the node features corresponding to the target to-be-explained sample; and determining a first distinguishing capability index of the target path node according to the node tree. And calculating a second distinguishing capability index of each target sample feature according to the first distinguishing capability index of the path node corresponding to the same sample feature. And calculating a third distinguishing capability index of all the sample features according to the second distinguishing capability index of the same sample feature in all the to-be-explained decision trees. In addition, the invention also provides a feature distinguishing capability quantification device, equipment and a storage medium.

Description

technical field [0001] The present invention relates to the field of big data technology, in particular to a quantification method, device, equipment and medium of feature distinguishing ability. Background technique [0002] The interpretability of a model refers to the extent to which model users can understand the model. If the user of the model cannot understand the relationship between the model's input and output, it is likely to make biased decisions and it is difficult to detect when the model is attacked. At present, models are gradually becoming popular in various industries. We urgently need to improve the interpretability of models to avoid unknown risks and meet regulatory needs. [0003] The isolation forest algorithm has the advantages of low computational complexity, easy processing of high-dimensional massive data, and distributed training, so it is widely used in the industry. However, this algorithm needs to randomly select features and randomly select t...

AI Technical Summary

Problems solved by technology

However, this algorithm needs to randomly select features and randomly select thresholds to divide data to build a large number of decision trees, resulting in low interpretability of features.

Images