Model establishing method and system based on feature importance

Abstract

The invention discloses a model establishing method and system based on feature importance. The model establishing method comprises the steps: S1, initializing feature data; S2, sampling the feature data to form a plurality of feature data combinations, and taking each feature data combination as a sub-model; S3, setting model parameters of each group of feature models, wherein each feature combination is provided with the same model parameter range; S4, training a model, and calculating the importance of each sub-model feature; S5, calculating the importance of the comprehensive weighted features by using all the sub-models; S6, sorting the feature importance to obtain a new importance sequence; and S7, modeling again according to the feature sequence. According to the model establishingmethod and system based on the feature importance, the relative sorting fluctuation of the feature importance obtained through calculation can be reduced, and the credibility is improved.

Description

technical field [0001] The invention belongs to the technical field of big data processing, and relates to a model building method, in particular to a feature importance-based model building method and system. Background technique [0002] In the field of big data risk control modeling, a variety of data sources are usually used, and due to the difference in data retrieval rate and the order of access time, the quality of data that can be used for risk control modeling is ultimately low. Faced with this kind of data situation, the performance of traditional algorithms is limited, and a large amount of data preprocessing is required, which cannot meet the risk control needs of the business. Therefore, choosing a model with higher performance is the primary goal, and XGBOOST is commonly used One of the algorithms, the content of this patent is based on the XGBOOST algorithm. [0003] The XGBOOST algorithm uses an ensemble learning scheme to build trees, and finally learns N t...

AI Technical Summary

Problems solved by technology

[0002] In the field of big data risk control modeling, there are usually many data sources used, and due to the difference in data retrieval rate and the order of access time, the quality of data that can be used for risk control modeling is ultimately low

Faced with this kind of data situation, the performance of traditional algorithms is limited, and a large amount of data preprocessing is required, which cannot meet the risk control needs of the business. Therefore, choosing a model with higher performance is the primary goal, and XGBOOST is commonly used One of the algorithms, the content of this patent is based on the XGBOOST algorithm

[0003] The XGBOOST algorithm uses an ensemble learning scheme to build trees, and finally learns N trees, whose feature importance is the ratio of the number of times they appear on tree nodes, because the test set overfitting and stability of the model need to be considered during the tree building process. For other factors, parameters need to be restricted, such as the maximum tree depth, node splitting conditions, etc., which will eventually lead to limited features used in the entire tree forest. Therefore, among a large number of modeled feature variables, most of the feature importance is zero. , the fact that most of the feature importance is zero is not conducive to the subsequent feature screening work, and the feature importance is zero does not mean that the feature is useless

Images