POI Representation Learning Based On Hybrid Model

The development of location-aware technologies such as mobile communications and sensing devices has led to an increase in the size and value of location data,which has become a strong support for the growing of Location-Based Service(LBS)and the treasure which academics and industry all concern.Similarly,POI(Point of Interest)has received more and more attention as a core element of location data.Focusing on two types of typical location data:trajectory data and check-in data,this paper has developed the following work.Firstly,the method of identifying stay point from trajectory data is studied.Naturally,the trajectory data has extremely high redundancy,and removing redundancy by the stay point identification technology is an inevitable choice for effectively utilizing trajectory data.Existing methods for identifying stay points have some shortcomings due to not considering time continuity or only considering one direction of time continuity.In this paper,a new method called Stay Point Identification based on Density(SPID)is proposed.SPID takes into account the spatial-temporal clustering of trajectory points,and the time directions and time continuity of trajectory points.The experimental results on Geolife dataset verify that SPID is better than the baseline methods.Secondly,the representation method of the POI in the stay point sequence and check-in data is studied.The sequence of stay points is similar in form and meaning to the check-in data,and they can be represented by POI.Therefore,simple and effective representation of POI is the primary of utilizing location data.Distributed representation is a universal and effective representation and the purpose of POI representation learning is to learn the distributed representation of POI,that is,to encode POI information into a continuous low-dimensional vector space.In this paper,in order to effectively learn the distributed representation of POI,we propose a hybrid model which mapping of the word vector model to the learning of POI type features and mapping the Network Representation Learning(NRL)model to the learning of POI location features to learn the POI representation of low-dimensional continuous vector forms.The hybrid model encodes POI information into a POI distributed representation in an unsupervised manner..We analyze the factors affecting the POI representation of the hybrid model by experiments,and analyzed the quality of the POI representation by two evaluation metrics.Finally,the effectiveness of the POI representation got from the hybrid model is analyzed by POI recommendation based on user-based collaborative filtering.