Feature combination method and device, medium and electronic equipment

Abstract

The invention provides a feature combination method and device, a medium and electronic equipment, and the method comprises the steps: taking a plurality of samples comprising a plurality of differentfeature domains as input of a neural network model, and obtaining an interpretation vector corresponding to each feature domain in the plurality of samples; marking the feature domain of which the interpretation vector meets a preset condition, the preset condition including that the absolute value of the interpretation vector is greater than a set threshold; sequentially selecting a preset number of combined feature domains as candidate combined feature domains according to the sequence of the marked frequencies of the feature domains from large to small; and selecting a target combined feature domain matched with the target model from the plurality of candidate combined feature domains. According to the method, by obtaining the interpretation vectors corresponding to the multiple different feature domains, the candidate combined feature domains having good influence on an output result of the neural network model are screened out, the selection range of the combined feature domainsis narrowed, the screening complexity of the combined feature domains is reduced, and the finally obtained combined feature domains have interpretability.

Description

technical field [0001] The present invention relates to the field of computer technology, in particular to a feature combination method, device, medium and electronic equipment. Background technique [0002] Automatic machine learning technology is gradually attracting attention in various fields because it can realize automatic data processing and modeling without deep mining of data. When using automatic machine learning technology for modeling, the selection of combined features has a significant impact on the accuracy of the neural network model. Combination features are formed by the intersection of multiple different feature domains. [0003] At present, when selecting combined features, one of the mutual information complete search strategy, brute force search strategy, and deep model dot product strategy is mostly used to realize feature domain interaction, so as to search out the feature combination required for the neural network model to predict the prediction ob...

AI Technical Summary

Problems solved by technology

However, the mutual information complete search strategy only uses the mutual information value as the screening condition of the combined features, which not only lacks generalization, but is only suitable for some classification tasks, and makes it difficult for the combined features to improve the prediction effect of the model.

However, the brute force search strategy needs to traverse too many combination features, which not only has a high failure rate, but also leads to a large time cost.

The combined features selected by the depth model dot product strategy are highly complex, difficult to be understood by humans, and not interpretable

Images