User Behavior Analysis And Prediction Baesd On Big Data In Cellular Network

In recent years,with the rapid development of information technology,more people are conducting various activities through different mobile terminals,such as listening to music,browsing news,watching videos,etc.,resulting in a large amount of data.The types and dimensions of data have grown and become more complex.Today,with the rapid development of mobile Internet,applications such as live video,short video,and feed stream generate a lot of valuable data.The ubiquitous business activities are constantly generating a large amount of information,and an era of information explosion has arrived.In this context,as more and more mobile Internet users are engaged in the data service camp,data services account for an increasing proportion.The growth of services in cellular networks is accompanied by the growth of traffic,generating a large aIlmount of user behavior data.Mining user behavior patterns from these data will effectively improve the load capacity of the cellular network and improve the performance of the cellular network.At present,there have been a lot of researches on user behavior analysis under cellular networks at home and abroad,but there are still two major problems:On the one hand,although there are many resear-ches on user behavior under cellular networks,most of them are for a certain dimension of user behavior.For example,starting from the dimension of mobility or the dimension of business usage,there is no comprehensive consideration,thus ignoring the correlation between them;on the other hand,most of the work is based on analysis,there is no deeper reason to explore the user behavior,that is,the exploration of the user's behavior is not enough.In view of the above two problems,this paper is based on the real 4G user behavior data of cellular network,and explores the laws and correlations of the three from the perspectives of data volume mode,business usage mode and mobility,and conducts user behavior based on these rules.Prediction.The main work and main contributions of this thesis are as follows:First,user data pattern mining based on cellular network.First,based on real data,the Spark big data processing platform is used to perform large-scale data cleaning,conversion,and aggregation of billions of levels,and extract the user's behavior track.Using the user's behavior trajectory,the usage rule of the data volume is further obtained,and most of the cellular network traffic used by a small number of users is found.Based on this law,the defined activation probability and the activation state transition probability,the data usage pattern is obtained and the user group division is performed,and the user is divided into "heavy users"and“ordinary users".Further,the differences in mobility and business usage between "heavy users" and "ordinary users" are compared.Finally,based on the business preferences of "heavy users",the non-negative matrix factorization algorithm is used to find the cause of "heavy users",that is,“heavy users”use some of the most traffic-oriented business categories,such as video,instant messaging,download,and E-commerce class.Second,business model mining and forecasting based on cellular networks.Firstly,the basic characteristics of the service mode are studied,such as the number of times the user accesses the base station in the time domain,the number of access service types,and so on.Secondly,the user service trajectory modeling is performed.Specifically,the user's business behavior is modeled as a vector,that is,the vector is used to represent the user's business trajectory.Using the unsupervised k-means clustering algorithm,the optimal number of clusters is determined according to the contour coefficients,and the user's business model is mined.From the results of the excavation,the differences between the modes are obvious.Some models have a high frequency of service usage during the day,while some models use the frequency of all services at a low frequency,and the peak frequencies of other business models appear at different times.Finally,based on the excavated user business model,a deep neural network model is built to predict the user's business model,and compared with classic models such as naive Bayes and logistic regression.The experimental results show that the model based on deep neural network can better predict the user's business model and perform better in terms of accuracy,recall rate and accuracy.From the performance of accuracy,the accuracy of deep neural network is 2.25%higher than logistic regression and 28.33%higher than Naive Bayes.Third,the user's mobile mode and business model correlation analysis and forecast.First explore the relationship between business models and mobile models.It mainly analyzes from the following two aspects:fir-st,analyzes the relationship between the number of visited base stations and service usage;second,the primary base station defines that the base station consumed by the user most of the time is defined as the primary base station of the user.Then,the main base station performs the extraction of the"comfort zone",and analyzes the service usage law of the"comfort zone".Secondly,for each user,identify the home and work place,calculate the distance between each user's home and work place,and compare the distance between urban and work places of different degrees of development,and dig out the degree of development between the city and the working distance.relationship.Finally,this paper uses a deep learning-based prediction algorithm to predict the click rate of various businesses under specific time and space conditions.Specifically,the feature is extracted according to the time and space information of each record,the uplink and downlink traffic,the uplink and downlink sending packets,and the like.The DeepFM algorithm based on neural network is used to estimate the traffic CTR,and compared with the logistic regression and factorization machine models.The experimental results show that DeepFM has higher prediction accuracy in this scenario,which is 4.07%higher than logistic regression.It is 2.26%higher than the factorization machine model.