Method and device for screening machine learning features

Abstract

The invention provides a method and a device for screening machine learning features. The method and the device relate to the field of machine learning models. The method comprises the steps of: performing preliminary screening on collected features by using a machine learning algorithm, so as to obtain a training feature set; carrying out model training by utilizing all features in the training feature set, and acquiring a performance result of total features of the training feature set; removing each feature in the training feature set separately, carrying out model training by utilizing the remaining features in the training feature set, and acquiring a performance result of the training feature set after the feature is removed; comparing the performance result of the total features and the performance result after the feature is removed, so as to obtain a performance result attenuation rate after the feature is removed; and determining the features meeting a preset condition to be a screened feature set according to the attenuation rate. The method and the device minimize the complexity of the model, thereby greatly reducing labor cost and time cost, and improving working efficiency.

Description

technical field [0001] The invention relates to the field of machine learning models, in particular to a method and device for screening machine learning features. Background technique [0002] At present, the modeling of financial-related models mainly collects a large number of features first, and uses machine learning algorithms to classify or regress large data. In order to easily obtain features that contribute to machine learning, it is necessary to sort out all features. In related technologies, artificially use business experience to select features related to it, and then comprehensively optimize the features in different dimensions such as interpretability and indicators, which can be applied to various and complex features that may affect financial-related models. There are many types. According to preliminary statistics, there are thousands or even tens of thousands of features that can be used for financial-related model training. Each step in the middle require...

AI Technical Summary

Problems solved by technology

[0002] At present, the modeling of financial-related models mainly collects a large number of features first, and uses machine learning algorithms to classify or regress large data. In order to easily obtain features that contribute to machine learning, it is necessary to sort out all features. In related technologies, artificially use business experience to select features related to it, and then comprehensively optimize the features in different dimensions such as interpretability and indicators, which can be applied to various and complex features that may affect financial-related models. There are many types. According to preliminary statistics, there are thousands or even tens of thousands of features that can be used for financial-related model training. Each step in the middle requires a lot of manual intervention, and the workload of manually screening features is very large. The lengthening of the modeling cycle has become the bottleneck of the entire modeling development cycle

Moreover, the selection of features directly affects the training effect of the model. Modelers need to have strong business experience, which greatly reduces work efficiency.

Images