An Optimization Method For Merchant Recommendation Based On Location Social Network

The widely used social networks based on location and the continuous expansion of their scale make the merchant recommendation system one of the hot applications.It is to recommend user an interested place.There are rich multi-source heterogeneous information in the social network based on location,such as social network relations,geographic location information including latitude and longitude,user's commentary of merchants,rating information,etc.Excavating these information can effectively improve the accuracy of the merchant recommendation system personally.The model based on social networks of existing models have used social relationship of social network to recommend merchants.However,the results of recent sociological studies suggest that this method may not be the most appropriate.Research points out that the rules that bring people together include:(1)habits or lifestyles;(2)attitudes;(3)tastes;(4)moral standards;(5)economic levels;and(6)people who are already known.Obviously,rules 3 and 6 are the mainstream factors considered by existing recommendation systems.But the biggest problem about the used social relationship of social network to recommend merchants is that although users are friends with each other,the difference of their lifestyle may be huge,so the result of a friend's recommendation may be completely different from what he hopes,and rule 1 may be the most intuitive,using social relationship with similar lifestyle for business recommendations may be more accurate.As the location-based social networks(LBSNs)becomes unprecedented popular,more and more people are eager to share their social activities that are recorded by visited locations(POI)online.These visiting location data are collected by LBSN companies as check-in records including user IDs,user check-in locations and the check-in timestamps.With the accumulation of check-in data on LBSN,it becomes possible to extract user's spatiotemporal preferences based on their historical checkin records.The shared check-in data from different users can also help big companies to provide more accurate location-based recommendation services.Currently,many LBSN companies like Foursquare,Yelp,Twitter,Facebook,etc.collect users' check-in data and recommend new POIs to users based on their preferences learned from the users' historical check-in data.More and more people rely on the location-based recommendation services to explore the POIs.Also,business companies can use POI recommendation techniques to predict user intent and thus improve their service quality.In the past few years,many techniques have been developed to extract users' preference from those check-in records generated by users.Many methods are based on probabilistic generative model(PGM).Those PGM based methods learn user's latent preference,such as latent spatial and latent topical preference,from users' check-in records.The posterior distribution for the latent spatial and topical variable can be inferred by the approximated methods,such as Markov chain Monte Carlo and Variational inference.Also,the matrix factorization based methods infer users' preference by factorizing the user-POI matrix into different meaningful latent feature matrices,the feature matrix represents user's preference.However,these methods can only learn user's static preference but can not capture user's dynamic interest that may change in different time period during a day.For example,people seldom visit a bar in the morning on weekdays as they must go to work,but they do at night when they get off work.The only exceptional PGM based method is USTTM which captures user's dynamic spatio-temporal topics from their historical check-in data by considering that a user may make choices at different time.Unfortunately,it captures user's time related POI preference from user's geographical check-in data denoted by the longitude and latitude without any semantic explanation.This may not explicit enough to capture users' preference for differen locations.This paper proposes an optimization method for merchant recommendation based on location social network about the above problems.It is divided into two parts,the improved model based on similarity calculation of social relationship and a new model for spatio-temporal topic preferences about geographic information.The original multi-source information fusion model CoSoLoRec contains content,social,and location information.In the modeling part of social relationship,a model based on similar social relationships of life styles is used for modeling,that is,in the social relationship modeling stage of CoSoLoRec,a user-based collaborative filtering method is used to predict the attractiveness of businesses to users.The LDA algorithm is used to introduce the content with lifestyle attributes to calculate the similarity between users,and the cosine similarity is used to measure with the improved Jaccard correlation coefficient.A novel scheme of model-ing user's preferences on spatiotemporal topics(UPOST scheme)is proposed about the modeling of geographical location information.In UPOST,we discovery user's temporal preference for different locations based on the semantic description of the location visited by users.We develop two algorithms,time division LDA(TDLDA)and time adaptive topic discovery(TATD).In TDLDA,we divide users' check-in data into different time segments according to the timestamps in every check-in records.Then Latent Dirichlet Allocation(LDA)is trained on each time segment with the corresponding semantic text for each check-in record.The learned topic representation of every time segment is used as the temporal preference of users for locations.To improve TDLDA,we provide TATD which considers the time variable and learns users' temporal preference from their check-in records.Compared with TDLDA that needs to train multiple LDA models,TATD could generate temporal topics within a single training model and produce more accurate results.The simulation results show that the improved model is superior to the methods in the current research work,and the results on the corresponding evaluation indicators(Precision and Recall)are improved.