User Behavior Prediction Using Cross-Platform Mobility Data

In recent years,the development of network and user services have greatly im-proved our lives,and recorded various categories of user mobility data,which reflect users' lifestyles,habitats,social status,even their characteristics.Users' future mobil-ity prediction using these data will help us exploring user behavior patterns,and enable appealing proactive experiences for services.Individual mobility prediction will offer essential business intelligence to advertising and marketing,bringing a more accurate recommendation system to them,while predicting future mobility patterns of crowds can help governments deal with public emergencies such as stampede prevention,and help them monitor the population flow among cities and within cities,providing new perspective for urban planning and the breaks of some infectious diseases.Moreover,short-term prediction will improve users' experience with intelligent personal assis-tants,help these platforms understanding users' contemporaneous intent,while long-term forecasting will provide inspiration to know users' long-term seasonal demand for certain commodity and hence find potential repeat buyers,and help developers to guide future investments.Several previous works have investigated the problem of user mobility prediction,trying to model user behavior patterns using various methods,e.g.,sequence analysis,Gaussion mixture model,matrix factorization,etc.However,the data recorded on one service platform can only reflect a mall part of users' lives,making the prediction partial and unconvincing.We hope to combine the heterogeneous datasets on different service platforms to describe a more accurate user profile and further help predicting his/her future mobility.There are many challenges for this task:when the users in different service systems could be linked,how can we deal with the data sparsity and the mis-alignment between datasets,when the users of different datasets are not linked,how can we leverage information from the other dataset,and when we only have anonymous mo-bility data,how can we exploit patterns of the crowds and help predicting the condition of a population.In this thesis,we use various real-world city-scale,even country-scale user mo-bility datasets to tackle the three problems mentioned above through specific and real situations.In specific,firstly,we combine the dataset of users' app usage and users'locations to achieve the long-term forecasting of their future app usage patterns.We propose an ubiquitous long-term forecasting model,which organically combines collab-orative method,time series analysis,and tensor decomposition.It explores the aggre-gation characteristics in the user,app,and context dimensions in dataset,and consider the stationarity,trend,and seasonality in temporal dimension simultaneously.Then,we utilize the check-in dataset of Weibo,anonymous bus records and taxi records to pre-dict users' future check-in locations.Due to that we can not linked these three datasets,we apply gravity model to learn the spatial influence between each POI venue to other venues and other regions from these datasets,and exploit the regularity and conformity from user mobility,to help modeling user behavior pattern and predicting user location.Finally,we utilize anonymous user check-in dataset to help predicting the health condi-tions in different regions of London area.We learn citizens' lifestyle information from their check-in dataset,and combine Gaussion mixture modeling and collaborative topic modeling to leverage the learned human lifestyles to improve the accuracy of citizens'chronic disease evolution rate prediction.We use various user mobility datasets and a chronic diseases dataset to evaluate our proposed models.The results indicate that our models outperforms the state-of-the-arts in both accuracy and efficiency.