Community Population Distribution And Individual Periodic Behavior Pattern Mining Based On Location Data

With the rapid development and maturity of positioning,wireless and Web technology,massive amount of movement data have been collected from various objects.As the volume of movement data grows,the impact of analyzing and mining these data becomes deeper in different areas such as echology,transpotation management,mobile communication management and surveillance of certain group of people.Analyzing and mining the objects' behavior pattern can be carried out through two different ways-by community or by individual.Considering from the macroscopic level,since individual behaviors are usually influenced by the community,the result of community behavior pattern mining can be a good guidance for mining and understanding individual behaviors.Considering from the microscopic level,since the community always consists of individuals that share similar behaviors and community behaviors are integrated representation of individual behaviors,individual behavior pattern mining is the solid base of community behavior pattern mining.Given the background and current situation of behavior pattern mining based on location data,it is obvious that although community and individual behavior mining are much different from each other,they share the same level of importance.As a result,this paper decides to explore and research from both macroscopic and microscopic level and visualize the research results.The main contribution of this paper can be divided into three parts.(1)Targeting a certain formal community which possesses comparatively large scale,relatively stable daily behavior and activity area with closure property to some extend,this paper proposes a population distribution prediction method based on multi-source data.First,we integrate different types of data,eliminate abnormal data,standardize data format and caculate population distribution with choosen sampling rate.Then we analyze the dominant factors that affect community behavior considering both the community characteristic and environment factors.And we then choose proper training set from historical data and updated data based on Pearson correlation coefficient.Using features and data above,we give a prediction result based on Bayes classifier and maximum a posteriori criterion.At last,we introduce an error-based learning method to adjust the prediction result.We use people on campus as an example in the experiments and analyze and predict the population distribution with data from some university in Shanghai.The experiment results stand for the above algorithms.(2)This paper proposes a periodic behavior mining algorithm that contains three stages-location data preprocessing,period detection and periodic behavior mining for the typical individual behavior pattern-periodic behavior.a)In data preprocessing,this paper proposes an abnormal location data adjustion method and a resampling mehthod based on layered mean value interpolation to deal with potential abnormal location data with large deviation and uneven sampled data.b)In period detection,cosidering that the object's location data may not be the same even it's at the same place,we first detect the area of important places based on kernel function and transform the spatial-temporal data into binary sequences according to different reference spots.Then,we introduce the algorithm that combines discrete fourier trasform and autocorrelation function to detect the periods in binary sequences.c)In periodic behavior mining process,we first use period along with a categorical distribution matrix to define a periodic behavior.Then,we introduce KL divergence as the distance measure between catogories and divide different periodic behaviors into different clusters by hierarchical clustering.The number of clusters is calculated through the change of representative error.Movement histories of primary and middle school students in Shanghai are used in the experiments and the mining results coinside with the students' actual behavior which in turn verifies the effectiveness of the periodic behavior detection algorithm.(3)The author developed an exhibition system for behavior pattern mining based on Java Web to visualize the results of above algorithms.Users can get to know the related technologies and results through interactive map,heatmap,dynamic line chart,selectable stacked bar chart and multifunction table with detailed information.She first designed and constructed the whole system based on MVC design mode.Then,she designed and implemented the main function modules and extensible database structure.Finally,she deployed the system on Tomcat,did the Web test and displayed the results in this paper.