Research And Application Of Telecom Package Recommendation Based On The FGCIN Model

With the rapid development of the information age,the mobile communications sector has also seen a sea change,with smartphone adoption across almost every industry and the number of users accessing the internet has grown wildly over the past few years.As the rise in mobile phone usage has leveled off,the three major operators in the country are now focusing their business priorities on customer retention and they need to introduce high-quality,more targeted packages to meet their customers’ needs.In order to retain existing users,telecom operators basically recommend packages through customer service outbound calls.However,the number of subscribers is so large that it takes a lot of manpower and resources to recommend packages to individual subscribers,so we need to interact with the characteristics of subscribers and packages to model and predict the probability of subscribers upgrading their packages.By targeting users who are likely to upgrade their packages and targeting them with outbound calls for package recommendations,the operator’s costs are significantly reduced.The end result of the recommendation of telecommunication packages is the user upgrade problem;while the study of click-through prediction is a click-through or not click-through problem,they are both binary classification problems.In this paper,it is of good practical importance to study the interaction of features in the click-through domain and apply it to the package recommendation problem in the telecoms domain,so as to predict the outcome of the upgrade performed by the user.However,most of the models tend to ignore the impact of different features on the model,and feature interactions can only occur at the elemental level,which prevents better learning of feature interactions;and most of the discrete features of the models are coded in a unique thermal way,leading to space wastage and other problems,and the telecommunication packages are often disorganized when training the model,requiring targeted feature processing.This paper proposes a new click-through rate prediction model based on the above-proposed problem and designs a telecom package recommendation system to apply to real life.The main work of this paper is as follows:(1)A click-through rate prediction model(Feature Generation and Compressed Interaction Network,FGCIN)based on feature importance fusion compressed interaction networks is proposed.The embedding layer of the model first uses Word2 vec technology to map discrete features into low-dimensional dense vectors,avoiding the conversion of discrete features into unique thermal coding,which has an impact on the model training;secondly,the important feature interactions are captured through the convolutional neural network FGCNN to generate new features to be combined with the original features and input to the feature interaction layer;then the compressed interaction network CIN is used in the feature interaction layer for The higher-order feature interactions occur at the vector level,allowing for better learning of the relationships between features and better interpretability;finally,the implicit feature interactions are performed by deep neural networks,allowing for more adequate feature interactions.(2)In response to the haphazard dataset of telecom packages,we integrate data processing with the model so that feature data processing can be performed automatically before each model training.In studying the data processing of telecom packages,four areas of work are included--missing data processing,anomaly data processing,data normalization,and positive and negative sample processing.In terms of missing data processing,the machine learning algorithms KNN and decision trees are used for continuous and discrete features respectively;in terms of anomalous data processing,the 3 criterion and box plots are used for the detection of normally and non-normally distributed features respectively;in terms of data normalization,we use minmax for normalization to reduce the impact of differences in feature scales on the model;in terms of positive and negative samples processing,the oversampling method is used to avoid the impact of imbalanced positive and negative samples on the final prediction.Finally,we conducted a large number of comparative,ablation,and hyperparametric experiments on the telecom package dataset and the public dataset Criteo to demonstrate the validity and reasonableness of our proposed FGCIN model compared to other mainstream models by evaluating the metric AUC.(3)Designed and implemented a telecom package recommendation system,deploying our proposed FGCIN model as well as other classical click-through prediction models into the system for use by business analysts in telecom companies.Business analysts predict the probability of users upgrading their packages through data import,new models,model training,model prediction,data export,and field management.For users with a high probability of upgrading,the operator’s customer service uses them as target users and recommends packages through outbound calls.Compared with the original recommendation method,this recommendation system effectively reduces the operator’s costs and is of practical significance.