Research On Point Of Interest Recommendation Algorithm In Location-based Mobile Social Network

The point of interest recommendation is to predict where the user wants to go next using the user's check-in data in the LBSNs.With the maturity of mobile positioning technology and the popularity of mobile terminals such as smart phones and smart wristbands,heterogeneous check-in data such as user's geographical location information,rating evaluation information can be more and more easily obtained in LBSNs.How to improve the recommendation performance with the multi-source heterogeneous check-in data of the user has become a research hotspot of POI recommendation algorithmTo a certain degree,the POI recommendation algorithm has achieved a lot so far.However,there are still some problems: Firstly,the POI recommendation is based only on single-dimensional check-in data,but ignores the correlation between the impact factors.For example,the recommendation algorithms for interest points based only on the time factor,or the recommendation algorithms based only on geographical factors,ignore the fact that the user's geographical position tends to change with time.But sequence in time also exists.Secondly,the rating data is very sparse in the LBSNs.The sparsity is as high as 98%,which is difficult to mine the user's real preference based on few rating data.It leads to bad recommendation performance.In view of the above two issues,the two factors of the four are combined according to their relevance.And the corresponding POI recommendation algorithm are proposed to improve the validity.According to the correlation between the geographical factors and time factors,temporal context and distance context information is introduced in the paper.In order to facilitate the modeling of temporal context information,this paper introduces the recurrent neural network model of deep learning filed into the POI recommendation algorithm.The first attempt was made using the recurrent neural network model.However,the experiments have shown that although this algorithm has certain effectiveness,it still has limitations.Therefore,some improvements are made in this paper to deal with the defects of the model such as serious vanishing gradient problem.The original model is replaced with the Gated Recurrent Unit model.The time-specific transition matrices is introduced in order to discretize continuous time factors.As well as the distance-specific transition matrices is also used.The POI recommendation algorithm based on the Gated Recurrent Unit model with the temporal and distance contexts is proposed.After contrast experiments,it is found that the new algorithm not only improves the effectiveness greatly,but also avoids the vanishing gradient problem.Otherwise,social relationships can alleviate the highly sparse problem of rating information in check-in data.Therefore,the rating information is linked with the social information.In addition,the real-valued conditional Restricted Boltzmann machine model has been proved as an effective solution to the recommendation because of its high accuracy.The paper integrates the friend information into the real-valued conditional Restricted Boltzmann machine model.The friend association is defined to improve the recommendation performance.The POI recommendation algorithm based on the real-valued conditional Restricted Boltzmann machine model with friend relationships and rating information is put forward.Finally,the verification experiment was conducted on the Gowalla dataset.The results show that the new algorithm still has good performance on the datasets with high sparsity compared with the traditional algorithm.