Research On Point-of-Interest Lifecycle Prediction With Mobile Sensing Data

With the rapid development of mobile social networks,smart device and human computer interaction related technologies,more and more inhabitants have been using Location based Services(LBS)for navigation,positioning and recommendation in their daily life.As a result,these LBS systems can record the evolutionary process of urban area with the mobile sensing data generated by different users.Along this line,a large amount of user generated content has been accumulating over time,which provide us new opportunities and challenges in exploring the hidden principles governing the development of cities.Meanwhile,with the development of big cities,a number of Point-of-Interests(POIs)emerge,grow,stablize for a period,then finally disappear.In other words,the POIs play important roles in providing diversified urban functionalities in big cities.Therefore,the evolutionary process of POIs can be deemed as the micro development of cities.In this dissertation,we introduce several recent research efforts on predictive analytics of the POI evolution based on mobile sensing data.Specifically,we first propose three research directions for predictive analytics of POI lifecycle,namely Evolution Modeling,Lifecycle Prediction and Evolutionary Trends Prediction.The research contributions of this dissertation can be summarized as follows.First,we study the POI lifecycle modeling approach.We propose the formulation methods of POI Lifecycle and POI Evolutionary Trends to uncover the POI evolutionary patterns.In particular,to handle the sparsity of POI data,we characterize the POI lifecycle with time series analysis.Besides,we formulate the POI Lifecycle Status by leveraging the temporal dependencies embedded in evolutionary snapshots.Meanwhile,we propose the concept of POI Popularity Joint Distribution with measure theory and develop the formulation methods for POI Evolutionary Trends.Finally,we conduct analytical experiments to investigate the POI evolutionary patterns with real-world POI data sets.Second,we propose a POI lifecycle detection approach based on multiple contextual information fusion.Specifically,we treat the POI lifecycle detection problem as a two-stage prediction problem: one can first predict the lifetime status of POIs in different time slices then detect the lifecycle of unknown POIs with Lifecycle Estimation Algorithm.Next,we integrate the geographical information and multi-scaled human mobility dynamics.In particular,to cope with the complex periodicity of human mobility behavior,we disaggregate the human mobility patterns at point,grid and territory granularities respectively.Finally,we conduct extensive experiments on real-world data sets.The experimental results validate our proposed approach in terms of feature effectiveness and lifetime status classification feasibility.Third,we extend kernel based vector machine and propose a data driven approach to inferring the POI lifetime status with multitask learning framework.This approach can effectively exploit the multi-sourced domain knowledge and adopt multitask learning as well as rare class classification methods to address the imbalanced classification problem.To be specific,we propose a novel classification method(MMKVM)taking into account the relation between classes and dependencies between time slices.In particular,to solve the optimization of MMKVM,we develop an effective learning algorithm by exploring the Laplacian regularizer,dual optimization and optimization decomposition.Finally,we perform extensive experiments using large-scale and real-world data sets.The experimental results validate the effectiveness of our proposed method in automatic inferring POI lifetime status.Fourth,we propose a CRF based approach to forecasting the rise and fall of volatile POIs(vPOI)over time,which can be applied to forecast the future trends of regional industries prosperity in a dynamic way.Specifically,we first treat the vPOI evolutionary trends prediction as a multi-variate regression problem.Next,to make the model more tractable without reducing the complex structure between different contextual variables,we develop a Dynamic Continuous CRF(DC-CRF)model to forecast the rise and fall of vPOIs.Moreover,by transforming the probability of DC-CRF into a Gaussian form,we develop an effective learning algorithm for parameter estimation.Finally,we conduct extensive experiments on real-world data from Google Maps.The evaluation results demonstrate that our proposed approach outperforms baseline algorithms with considerable margins.To sum up,in this dissertation,we propose several data driven approaches to address the POI lifecycle prediction problem arose in modern urban development.We validate these methods on large scale real-world data sets,which can be beneficial for the study on urban development theory as well as the urban computing applications.