Personalized Poi Recommendation Based On Geographical Influnce And Visual Contents In LBSN

With the popularity of cellphones and the widespread application of global positioning systems,location-based social networks(LBSNs)have become one of the most popular way for people to share their check-ins at point-of-interest(POI).Meanwhile,with the rapid improvement of information technology,users often suffer from information overloading problem,which means finding locations that satisfy users' interest in LBSN is hard.This problem promotes the emergence of POI recommendation algorithms for providing high quality location recommendations.Based on the fact that the geographical clustering phenomenon and sparsity problem in users' check-ins data,it is crucial to design models based on the features of the data in LBSN.Previous works were mainly on the study of utilizing social relationships,time factors,geographical influence and contents information for designing POI recommendation algorithms.Related works demonstrate that social relationships and time factors have their limits,while geographical influence and contents information can promote the performance of the recommendation algorithms.However,the existed research works on these two factors were separate and their performance were not ideal enough.Algorithms will suffer from data sparsity problem when only considering geographical influence.Similarly,algorithms will generate some locations which will not be visited by users for the reason of geographical locations limits when only taking contents information into account.Comparing to text,images contain more information which can simultaneously reflect users' preference and locations' characteristic.The combination of geographical influence and visual contents for designing recommendation algorithms to boost the performance of the recommendation algorithms is an urgent problem to be solved.In this thesis we study the POI recommendation algorithms in LBSN,and in view of the weakness of previous works,we consider combining geographical factor and visual contents for jointly simulating geographical clustering phenomenon and relieving data sparsity problem.We proposed a merge matrix factorization model based on these two factors and give a learning algorithm for this model.The main work includes:Firstly,based on a real dataset in LBSN,we filled the missing images by a webpage crawler we designed.After cleaning the dataset,we analyzed the statistics of this dataset to show the validity of this dataset.Furthermore,we proposed a POI recommendation model based on geographical influence and visual contents.In the geographical influence modeling part,we proposed a clustering based location region model to properly segment geographical locations.Base on this location region model,we proposed a latent factor model based on location-region jointly appearance matrix to simulate geographical clustering phenomenon in users' activity.While in the visual modeling part,based on images visual contents features of users and locations,we designed a visual distance computing model.And in the model merging part,we proposed a negative samples weight filling method based on the visual distance between users and locations which can relive data sparsity problem.At last,we proposed an algorithm to learn the parameters of this model.Finally,based on a real check-ins dataset,we carried out experiments to verify the validity of proposed model.Through the comparison and analysis of the experimental results with related algorithms we can get the following conclusions: The proposed POI recommendation algorithm based on geographical influence and visual contents can provide high quality location recommendation.This algorithm obtaining approximately 7% precision and recall improvement over the contrast algorithms,and the jointly usage of geographical factor and visual contents can promote the performance of the recommendation algorithm.