Research On Users And Their Behavior In Cross-device Identification

With the widespread use of a variety of intelligent electronic devices in people's lives,such as smart phone,laptop,tablet and other mobile smart devices,people are increasingly inclined to through different intelligent devices to complete the same task.However,when users switch back and forth between different smart devices to complete the task,their identification information will become very vague,so it's difficult to track their traces.Advertisers want to find users who use these devices instead of the device itself,so that they can do some precise marketing for the tasks that users have to accomplish.Therefore,in recent years,cross-device identification users become a popular research theme gradually,it is great significance on improving the value of advertising and enhancing the user experience.In recent years,many ordinary families or individuals have a number of intelligent devices that can be connected to the Internet,with the advent of the Internet of Things,smart devices will show explosive growth.Advertising companies always want to seamlessly connect to the consumer behind the device,not the device itself.The traditional approach is to use certain deterministic features,such as phone numbers,ID numbers,e-mail,etc,these features require users to provide initiative,similar to our daily use of the login account.When a user switches between different smart devices,a user can be uniquely identified based on deterministic characteristics.However,for privacy and security considerations,users are likely to refuse to provide their private information,which to cross-device identify users with great difficulties and challenges.In the view of the shortcomings of traditional cross-device identification technology,some scholars have tried different improvements in traditional methods in recent years.Most of the research is based on the analysis of user behavior consistency.With the rise of machine learning,some scholars put forward probabilistic machine learning methods to predict the probability that users have a certain intelligent device,which greatly improves the accuracy of cross-device identification.However,the size ofthe behavioral data of cross-device users is usually very large and very sparse,and some of the existing studies are often faced with data size and time consumption.In this paper,with the increasing demand for cross-device identification users,we deal with a large number of users' cross-device behavior data,and analyze the advantages and disadvantages of the existing machine learning methods.We discuss the problem of time consumption and data sparseness,propose a improved FFM model,which solves the problem of cross-device identification user.The contribution is as follows:1)Analysis of user behaviors and feature space: A detailed analysis of cross-device user behavior,and a more complete pre-processing of the data,through the analysis of the relationship among the characteristics,we generate a more complete feature.While the missing values of the data processing,carried out One-Hot coding.We also do data standardization and other processing,so that the prediction results better.2)Comparison of different algorithms: Eight machine learning models that can be used to solve cross-device identification user problems are compared,and their performance on different data sets is compared.At the same time,the time consumption of different models is discussed.We prove that XGBoost is a better model to solve the problem,and explore the optimal parameter selection problem of XGBoost model.3)An improved FFM model is proposed.For the sparseness of large-scale data,the FFM model can be better processed.On the basis of it,considering cross-device user's behavior of the region characteristics,proposed an improved FFM model.The model can solve the large scale sparseness problem of data,and the experiments show that our model has a better effect and more stable characteristics in the results of eleven models.At the same time,we explore the optimization problem of parameter selection.