Research And Application Of Telecom Package Upgrade Prediction Model Based On Deep Learning

With the rapid development of mobile Internet,smart phones have been popularized rapidly,and the scale of Internet users is also increasing year by year.Smart phone Internet access is inseparable from the support of telecom operators.In order to better meet the needs of users and provide satisfactory services,operators will launch some new packages and activities,introduce them to users through outbound customer service,and guide them to upgrade packages.The target of the traditional outbound call method is very broad,resulting in the profit and the time it takes are often not worth the loss compared with the energy.Therefore,finding an effective way to model the user’s basic information,package information,user consumption and other information,so as to predict whether the user is willing to upgrade the package,so as to accurately call out those customers who are willing to upgrade the package,which is very important for operators to improve service quality and user experience.The click-through rate prediction research is to obtain the probability of users clicking on the product or whether they click on it under a given scenario and product,which is very similar to the purpose of Telecom package upgrade prediction.This paper studies the feature interaction modeling in the hit rate prediction model,and applies it to the telecom package upgrade prediction,which is of great significance in theoretical research and practical application.Although great progress has been made in the research of hit rate prediction models based on deep learning,most models either model feature interaction under predefined order,or they can not explicitly model feature interaction,so they do not have good interpretability.In addition,the existing processing methods for continuous features in data sets have some problems,such as insufficient utilization of continuous features,low capacity,and no support for end-to-end training.Based on the above problems,this paper proposes two hit rate prediction models and applies them to telecom package upgrade prediction.The main work of this paper is as follows:(1)Based on attention network and adaptive factorization network,a prediction model MAFN(Multi-head Attention and Adaptive Factorization Network)integrating explicit and implicit feature interaction is proposed.Firstly,the model uses the multi head self attention layer with residual connection to capture the explicit feature interaction.By stacking multiple such layers,we can learn the explicit high-order feature interaction.In order to make up for the deficiency that the multi head attention module can only learn the feature interaction of a specified order,we use an adaptive factorization network with a logarithmic transformation layer to learn the feature interaction of any order.In addition,in order to learn more nonlinear characteristics,the residual network is integrated to enhance the generalization of the model.Compared with deep neural network,it can fit the information difference between input and output better.Finally,the outputs of the three sub modules are fused and spliced through the parallel structure,and the final prediction results are obtained.The experimental results of the model on three datasets show that it has achieved better results than the current popular similar models in terms of AUC and logloss.(2)Combined with an automatic discretization method of continuous features and an improved calculation method of self attention---decoupling self attention,a second prediction model DEMAN(Disentangled Multi-head Attention Net Work)is proposed.In order to obtain the unique embedded vector representation of continuous features,it is sent to an autodis module,which supports end-to-end training and has high model capacity.In the feature interaction layer,the unary term and second-order interaction term in self attention are decoupled to avoid the gradient interference between the two terms in back propagation.Finally,the high-order feature interaction is obtained by stacking multiple multi head decoupling self attention layers with residual connection,and the deep neural network is integrated to improve the prediction performance of the model.The model is tested on two datasets with continuous features.The comparative experiments between the models show that its performance is better than that of similar models.The comparative experiments of different processing methods of continuous features show that the discretization method we adopt is effective.(3)Based on the prediction algorithm proposed in this paper,a telecom package upgrade prediction system is designed and implemented.The system is oriented to business analysts and integrates the functions of model training,model prediction and data import.After the training of actual outbound call data,the prediction results of the model are truly applied to customer service outbound calls,which effectively improves the success rate of user upgrade package.