A Research Of Trajectory Clustering Based On Metric Learning

As more and more products and services based on location are applied,the volume of trajectory data is increasing day by day and trajectory data has become a common data type.Faced with a large amount of trajectory data,how to effectively mine and analyze it has become an important issue.Trajectory data has many basic features that differ from other data types: 1)Temporality: Trajectories are site coordinate information collected by sensors or positioning devices at the same or different time intervals,2)spatial variability: Trajectory data is differently distributed in the regional space,3)tight coupling of time and space: trajectory data has both temporal and spatial characteristics,and the relationship between time and space is extremely tight,4)complex structure:usually the length and sampling rate between trajectory and trajectory are much different.In this paper,we use spatio-temporal trajectory data as our study object and figure out effective clustering and anomaly detection algorithms.In all trajectory mining techniques,trajectory clustering and abnormal trajectory detection played important roles.Through the clustering of trajectory data and the detection of abnormal trajectories,it is possible to find similar or abnormal motion patterns existing in trajectories,thereby providing effective information for practical applications such as traffic flow monitoring,hot spot discovery and user behavior pattern discovery.This paper mainly contains two research contents.Due to the existing techniques simply migrate the time series algorithm to the trajectory data or only consider the similarity in geometry space,and have not proposed a method that can well combine the trajectory spatiotemporal information to characterize trajectories and measure distance among them.This paper firstly proposes a trajectory clustering algorithm based on metric learning,which can effectively extract the similarity information of trajectories both in time and space,and iteratively optimizes trajectory clustering results and metric function between trajectories in a similar space.In this way,we can eventually obtain trajectory feature vectors,trajectory distance function and also trajectory clustering results.Secondly,for the global trajectory outlier detection problem,this paper proposes an abnormal trajectory detection algorithm based on density clustering.The detection algorithm first extracts the contextual similarity information of trajectories,constructing the trajectory feature representation.Then discovers the global abnormal trajectory point through the density clustering idea,and finally proposes an anomaly scoring factor based on kernel density function.With this abnormal factor,the global anomaly detected could be sorted and global Top-N anomalies could successfully be found.