Research And Application Of App Usage Behavior Based On Heterogeneous Network Representation Learning

In the era of mobile Internet smart phones,people's lives have been greatly facilitated,and network users have put forward higher requirements for Internet services.Analyzing and predicting users'App click behavior will not only help to improve the quality of network service,but also help users quickly start the required App and save time and cost.Starting with DPI traffic data of operators,this paper discusses the research and analysis methods of mobile user behavior.The main work of this paper is as follows:(1)Based on DPI traffic of operators,this paper constructs network fingerprint data to study user behavior.A user-App-category heterogeneous information network is constructed from fingerprint data,and a HINE algorithm for learning the relationship vectors between heterogeneous network nodes and metapaths is proposed.HINE algorithm includes random walk and neural network learning.The three random walk strategies proposed in this paper can effectively represent the behavior preference between users and Apps and the extended relationship between different subnetworks.Through visual analysis of App vectors and experiments on open datasets,it is proved that HINE algorithm can effectively lean the feature vector representation of network nodes.(2)On the basis of network representation learning,this paper designs App click behavior prediction models for mobile users,which are MLP model and LSTM+Attention attention mechanism model.MLP directly uses the network representation to learn the user and App vector representation,and uses multi-level feature combination learning to predict the user's probability of using this App at this time.LSTM +Attention model aims at the sequence of user's historical behavior.LSTM model is used to model and learn user's historical behavior.Attention is responsible for stimulating the current candidate App's historical behavior to get users'instant interest,which greatly improves the prediction effect of the model.Finally,the AUC value of HINE+Attention model is 0.8715,which is 0.37?1.10%higher than the baseline work.This paper presents the concept and method of network fingerprint to deal with mobile users'App usage behavior data,and the research work of App click behavior prediction based on heterogeneous network representation learning using network fingerprint data,which provides a good idea for the current research and application of Internet users,behavior prediction.