Research On POI Recommendation Method Based On Multi-factor Influences

In recent years,the interaction between people and the Internet has become increasingly frequent and close.With the rapid development of various mobile phone applications and network media,Internet users have to search for their interesting content from mass data every day,so recommendation system has begun to play a role in the Internet.This thesis mainly focuses on the application areas of the point of interests recommendation.The system focuses on helping mobile users explore new areas.The purpose is to predict the possible behavior types of users at the current time by using user preferences and the characteristics of the point of interests,so as to recommend the optimal list of search points for users to choose.This thesis mainly includes the following aspects:In view of the current situation of information diversification in the field of recommendation,this thesis aims to build an end-to-end recommendation model by integrating a variety of different data types.In order to solve the problem of multi-degree relationship between users and point of interests,this thesis uses graph convolution technology to learn the features of interactive graph network between users and point of interests.In order to solve the problem of text data modeling,this thesis constructs the corresponding vector expression of the brief text of point of interests and user comment information respectively by means of sentence vector modeling based on attention mechanism and grammar analysis.In order to capture the change of user's interests effectively,this thesis divides the user's historical behavior into long-term interest sequence and short-term interest sequence,and then uses recurrent neural network and attention mechanism to model the sequence data,and converts user's behavior into a reasonable vector expression.Finally,the points of interest are recommended to the users by integrating the above three features and some artificial discrete features,statistical features,and spatio-temporal features through an end-to-end network.In addition,in order to solve the information migration problem caused by the decoupling between the recall phase and the sequencing phase in the recommendation process,this thesis also introduces a candidate set recall strategy based on vector and text features,which effectively combines the recall phase and the sequencing phase together.Firstly,the candidate list is recalled by the similarity calculation of the user vector and the vector of point of interest.Then,the candidate list is further fine-tuned by the text comment polarity,and then the time statistics information is used to constrain the result.Finally,the candidate set of the points of interests needed for recommendation is obtainedThis thesis verifies the proposed method in detail by testing on a real data set.Compared with some baseline models mentioned in this thesis,the experimental results show that the proposed algorithm can achieve certain results in the accuracy of recommendation results and the recall of candidate sets.