| With the rapid development of mobile devices and Web2.0,the Location-Based Social Networks gradually become popularized in the our daily life.For now,TB-level data is generated by the mainstream social applications every day.These data were recorded in the form check-ins.Researchers have proposed various location recommendation models based on these data,but these models do not perform well in dealing the data sparsity and cold start problem,the time efficiency and recommendation accuracy are low at the meantime.This paper proposes a POI mining method based on graph embedding that corporate the side information of POI and POI itself to generate POI Embedding.Then we integrate the POI Embedding to a location recommendation model based on the Encoder-Decoder framework,and we finally construct the location recommendation model.The main work of this paper is combined by three tasks:(1)POI mining based on Graph Embedding.Firstly,we extracted user move sequences and POI from the history check-ins data of users,then we built a POI-POI directed graph according to the sequences and POI.Secondly,we reconstructed the user move sequences based on Random Walk.Lastly,the POI Embeddings were generated.Based on the model intorducted above,we proposed an advanced POI Embedding generating model that incorporated with POI side information and the weighted mechasim which means the importance of different side information is different.(2)Location recommendation based on Encoder-Decoder.We utilized a global encoder to capture the long term preference of user and a local encoder to mining the short-term preference of user.Then we incorporate these two encoder to gengerate the User Embeddings which contain abundant information and can reflect the preferences of user precisely.Based on these work,we put the POI Embeddings which were generated in task(1)and User Embeddings into the decoder to get the location recommendation results relatively precise.(3)The experiments of model mentioned above.We do experiments on a large real Datasets Foursquare which exploit the POI Embedding generating model in task(1)and the User Embedding generateing model based on Encoder-Decoder.We also design different experiments in different scenarios and the results show that,the location recommendation of this paper perform well on Accuuacy@k,and in the experiments of different scenarios,the precision and efficiency of our model were the best comprehensively. |
