Abstract: Customer classification has always been one of the most important issues in customer relationship management (CRM). Therefore, it is very important to select key features of customers. In the era of big data, unbalanced classification distribution, high dimension, and a large number of samples without label have made this more complex and become a complex systemic decision issue. In order to address this issue, this study proposes the semi-supervised key feature selection model of customers based on hall for workshop of meta-synthetic engineering (SFS-HWME). The model invites five experts in related fields to identify research difficulties, find alternatives through qualitative analysis, obtain a total solution through comprehensive integration and get a quantitative analysis model. The quantitative analysis model uses semi-supervised learning (SSL). Firstly, it uses the data set L with category tags to train the Adaboost integration model to predict the categories of samples in the data set U with unclassified tags; secondly, the data set U is clustered by the self-organization map (SOM) algorithm and the samples are selectively tagged; thirdly, these samples are added to the data set L along with the tagged category tags; finally, the re-sampling technique is used to balance the class distribution of the new training set L, and the group method of data handling (GMDH) deep learning network is trained to pick out the optimal feature subset. The research invites 5 experts to select the most reasonable features. The empirical analysis on four customer classification data sets shows that the proposed SFS-HWME model has better key feature selection performance than some existing models.

Key words: hall for workshop of meta-synthetic engineering, customer classification, feature selection, semi-supervised learning, GMDH, re-sampling