Research On 5G Potential Users Prediction For Imbalanced Data

Nowadays,the telecommunications industry market is highly saturated,Competition between operators is increasing accompanied by the large-scale implementation of the transfer network,the rise of 5G makes the fiercely competitive operators have a new turn of events.How to seize the opportunity of the rise of 5G,identifying users' demand for 5G based on existing 5G user information timely,as well as providing related services pointedly are problem requiring special attention by operators.In this context,this paper takes data mining technology as the main line,analyzing and processing relevant information of existing 5G users,establishing a prediction model for potential users from the perspective of class imbalance problem.The main research work of this paper is as follows:(1)Based on the understanding of the business,conducting describe and multilayer division for the characteristic dimension of the data,a series of preprocessing operations such as cleaning data,removing unique attributes,transforming data,data extraction and so on are carried out on the data,making feature construction and multiple feature selection for preprocessed data.Not only rich data attributes,but also select data features that are highly correlated with target variables and have low redundancy between them applied to model training.(2)Establishing the classification prediction model based on 5G potential user prediction problems,Specifically including logical regression model and decision tree model,XGBoost model and Light GBM model based on Boosting strategy.Comparing and analyzing the performance of each model on accuracy rate,recall rate,F1 value,AUC value,proving the predicted effect of 5G potential users of the latter three belonging the ensemble learning model better than the former two belonging to a single machine learning model.(3)To make up for ensemble learning model which with the problem of data equalization addresses missing issues,this paper propose to make SE(SMOTE?ENN)method which including SMOTE sampling method and noise reduction method ENN integrate into an integrated learning model,what we respectively call SE?RF model ?SE?XGBoost model and SE?Light GBM model,and comparing and analyzing single SMOTE oversampling method and ENN subsampling method impact on user data imbalance problems.The results show that the SE method improves the prediction performance of the model more obviously than the two methods.(4)In order to improve the classification ability and generalization performance of single integrated learning model,this paper will integrate heterogeneous SE?RF,SE?XGBoost and SE?Light GBM three models based on Voting customize weighted soft voting strategies.By making a comparison,it is not difficult to find this method have more predictive performance than hard voting strategy based on the rule that the minority is subordinate to the majority and soft voting strategy based on the default parameter to the weight value on users of 5G.