Modeling of Human Movement Behavioral Knowledge from GPS Traces for Categorizing Mobile Users Sourse: Www 2017 Advisor: Jia-Ling Koh Speaker: Hsiu-Yi,Chu Date: 2017/11/7
Outline Introduction Method Experiment Conclusion
Introduction Question Can we map the knowledge of one known region to another unknown(target) region and use this knowledge to categorize the users in the target region?
Introduction Goal Labeled GPS log Unlabeled GPS log Knowledge User category label
Outline Introduction Method Experiment Conclusion
Method
Method User Trajectory Segment <S[], W[], Traj_Win[]> S[]: list of stay points, s = <lat, lon, Geotagg> W[]: list of waiting points, w = <lat,lon> Traj_Win[]: {S1, (x1,x2), (x2,y2), … ,S2}
Method User Trajectory Segment
Method Semantic Stay Point Taxonomy(SSPTaxonomy) SSPTaxonomy:<N, Nc, W> N: place type of the Taxonomy Nc: associated code of the node place W: aggregated footprints of user
Method Semantic Stay Point Taxonomy(SSPTaxonomy)
Method User-Trace Summary(UTS) NB = <G,Θ>,G=<V,E> v1i: (Nc,ti) Nc: associated code of the node place ti: temporal value of the node ei: dependences between the vertices
Method User-Trace Summary(UTS) Bayesian Network
Method Θx5|Pax5 = 0.46
Method Θx4|Pax4*Θx5|Pax5*Θx2|Pax2 = 0.6*0.46*0.98=0.27048
Method Temporal Common Sub-sequence, (TempCS)clustering algorithm Similarity measure(Bhattacharyya distance): DB(X4, X5)=-ln{[X4(0)X5(0)]1/2+[X4(1)X5(1)]1/2}= -ln{[0.4*(0.4*0.32+0.6*0.54)]1/2+[0.6*(0.4*0.68+0.6*0.46)]1/2}
Method Temporal Common Sub-sequence, (TempCS)clustering algorithm NB1:X4X3X5X1X2 NB2:X4X1X6X5X2 Common stay points(Lc):X4X5X2X1 Common Sub-sequence(Ls):X4X5X2
Method Similarity between NB1and NB2: SimSequence(NB1,NB2)= 3/4[DB(X4, X5)+ DB(X5, X2)]
Method User Categorization Classification task: PVu = {p1,p2,…,pi} i: user-category pi: probability of the user u in category i
Method User Categorization Feature f1: visit in types of places f2: Speed of movement or transportation mode f3: User Movement
Method User Categorization Bayesian network When independent Weighting each of feature
Method Transfer Learning
Method Transfer Learning Extract the parent’s code cp of a node c. Node c has n sibling, append n+1 along with the parent’s code cpn+1. Check whether the same place-type in and assign the same code if present. Generate Get the common taxonomy
Outline Introduction Method Experiment Conclusion
Experiment Dataset
Experiment Accuracy of User-Classification
Experiment
Outline Introduction Method Experiment Conclusion
Conclusion Address the user categorization problem from the GPS traces of the users. Propose a framework to model individual’s movement patterns. Transfer knowledge base from one city domain to another unknown city.