Location Category Prediction Based On Embedding Learning

With the wide spread of GPS-enabled smart mobile devices and the rapid development of online social networks,location-based social media has emerged.In such social networks,users can share their locations with friends through check-ins.These applications continue to explore a user's points-of-interest(POIs).Exploring the semantic meanings of these POIs can help operators infer a user's interest,predict the target's next movement,and improve the performance of product recommendation services.In addition,it can serve as the basis of many urban computing applications,e.g.,personalized route recommendation and user trajectory clustering.According to statistics,about 30%of POIs in the check-in data lack corresponding semantic category information,and the existing semantic categories are mainly manually annotated,which requires a great deal of time and labor costs.For location category prediction,current methods mainly regard it as a multi-classification problem.These methods firstly extract features(e.g.,visiting frequency,stay duration and distribution of check-in time)based on the multi-dimensional attributes of check-in records.Subsequently,they use the extracted features to infer the category labels of POIs,which often employs traditional classifiers,such as support vector machines(SVMs)and logistic regression,or recourses to neural networks,such as multi-layer perceptrons(MLPs).Nevertheless,the performance of these classifiers depends heavily on the way of feature selection.As there is no principled method to follow,it can be extremely daunting and time-consuming to determine a set of proper features.In this paper,we seek for a fire-new strategy to model the check-ins in location-based social networks.Inspired by the idea of word embedding in text mining,we propose a Location Category Embedding(LCE)model,which projects the POIs and their associated category labels into the same low-dimensional latent vector space.The LCE model has two advantages:(1)to capture the factors that might affect users'moving behavior,it considers sequential pattern,personal preference,and temporal influence,and further models the connection between the POIs and these three factors;(2)it minimizes the distances between the POIs and their category labels in the latent vector space,resulting in the POIs with the same category label are clustered closely around the category.At the forecasting stage,given an unlabeled POI,we calculate the distance between the target POI and each possible category label in the latent space,and then select the one with the minimum distance as the prediction result.We conduct extensive experiments on real check-in datasets,and compare the performance of the proposed LCE with several state-of-the-art baselines in the task of category label prediction.Experimental results prove the effectiveness of our LCE.