Research On Point-of-interest Recommendation Based On Personalized Geographical Influence

With the advent of Big Data era,it has become a challenging research issue to ob-tain effective information from massive data.Among various information extraction ser-vices,the point-of-interest(POI)recommendation service mines users' geographical pref-erences and predicts users' interested locations based on users' check-in information,which has been widely concerned by industry and academia for its outstanding business value.Most advanced POI recommendation models have combined with spatial data through ex-plicitly regarding geographical relevance between two POIs as a co-occurrent probability.However,such works consider spatial influence as only POI-related,poorly capturing the user-dependence of geographical influence.According to the different classification of rec-ommendation tasks,this paper puts forward general and next POI recommendation models based on personalized geographical influence respectively.Specific contributions mainly include the following two aspects:(1)For general recommendation tasks,the paper proposes a personalized geographical influence modeling for general point-of-interest recommendation(Gen-PGIM).First-ly,Gen-PGIM improves the neural collaborative filtering framework for recommen-dation diversity.Combined with the spatial data of POIs,Gen-PGIM then establishes personalized geographical relevance between POIs from three aspects:global dis-tance tolerance,local distance tolerance and the distance between POIs.Extensive experiments on three real-world datasets show that Gen-PGIM outperforms previous geographical recommendation models in precision,recall and diversity of recommen-dation.(2)For sequence recommendation tasks,the paper proposes a personalized geographical influence modeling for next point-of-interest recommendation(Next-PGIM).Next-PGIM combines the user's check-in sequence and corresponding distance sequence to capture the user's sequence preference and geographical preference for the next moment.Then,due to the temporal and semantic irregularity of the user's check-in behavior,the time and semantic sensitive sequence(geographical)preference is es-tablished.Finally linear and nonlinear fusion methods are used to combine the two preferences.Experiments on three real datasets show the effectiveness of Next-PGIM in improving recall and mean reciprocal ranking.