A Typical Point-of-interest Recommendation Approach Based On Geo-Social-Comment Relationships

The existing POI recommendation methods mainly use the geographic information of the POI and the social relationship of the users to improve the recommendation efficiency.However,the influence of the users' comments associated to POIs is neglected even though the comments are very important for users to make a decision.Furthermore,the POIs in recommendation list are usually similar to each other which makes them losing typicality and diversity.To deal with the problem mentioned above,this paper combines the POI's geographic information,user social relations,and comments to build a new model for measuring the relevancy of the POIs,which is called Geo-Social-Comment relationship model.Further,a new method to measure the similarity of comment text is proposed.According to the Geo-Social-Comment correlation between POIs,the POI clustering method based on spectral clustering and the typicality selection method based on probability density-based are proposed to pick a representative POI from each cluster.Next,this paper proposes a probability factor model which combines Poisson distribution with matrix decomposition to fit the frequency matrix of user's visiting POIs,and then the typical POIs selected in each cluster are ranked according to the user preferences.Experimental results demonstrate that it is more reasonable to evaluate the relevancy of POI by using the model of Geo-Social-Comment,and the recommended results have achieved better results in diversity and accuracy.In addition,for the traditional text similarity evaluation method,the text vector dimension is large,the data is sparse,and the word semantics and text grammar are not considered.We proposes two user comment text similarity measuring methods.(1)WordNet & string short text similarity-based method(WN-SS),which combines the similarity between strings and the semantic similarity between words,fully considering the semantic information,and to a certain extent solved the problem of misspelling words.(2)LU-CNN,which integrates word vector and convolutional neural network,solves the problem of missing semantic grammar information in traditional model by fully extracting the text depth features.Experiment results demonstrate that the two algorithms are more approach to the results of user labeling and have achieved higher accuracy.The paper includes 24 figures,16 tables and 55 references.