Research On Time-aware Hybrid Point-of-interest Recommendation Algorithm

With the development of deep learning,the field of artificial intelligence has made revolutionary progress,the research in various fields has also developed rapidly.Among all of them,the recommendation system applies the deep neural network model,which has been further innovated and widely accepted by scholars.In recent years,portable smart devices have become more and more popular,people's interaction with the Internet has become more frequent,and the accuracy of location tracking and positioning has been continuously enhanced,making location-based services tend to be diverse.Users can share their own location and comments in social networks,which makes point-of-interest recommendation a hot task in academia and industry.Point-of-interest recommendation not only requires the algorithm model to extract the effective information in the position from the massive data,but also needs to be able to cross-combine the position information with other heterogeneous information to accurately make personalized recommendations.This paper mainly studied the application areas of point-of-interest recommendation,which focuses on helping mobile users to explore new areas and predicts the behaviors that users may make at this time through various factors such as user preferences and places' characteristics.Then,recommending a list of best places for the user to select.This article includes the following aspects of work:Firstly,the overall architecture of this paper is modeled by the wide&deep model.In the deep side,three features are constructed,which are distributed expression of users,distributed expression of points of interest and contextual distributed expression with enhanced time feature.The latter two are the key features of this article.For the distributed expression of points of interest,this paper used a large number of comment information,applied word embedding technology to transform the spliced text into feature vectors,and obtained distributed expression of points of interest through convolutional neural network.Convolutional neural networks can acquire features such as n-grams,and embedding expressions of point of interests can effectively extract a large number of potential features.For the context-distributed expression with enhanced time features,since the checkins presented in the form of a sequence,this paper input the distributed expression vectors of points of interest into the long-term and short-term memory network which can simulate the sequence features well.Long-term sequences and short-term sequences were constructed to simulate behavior habits and interest changes of users,and then spliced them to obtain enhanced contextual distributed expression.This paper also briefly introduced the feature engineering of the wide side,and further improved the generalization performance of the model through discrete features,statistical features and spatiotemporal features.Finally,in this paper,the deep side and the wide side were unifiedly trained,and the obtained probability values were sorted,the best pre-judgment set obtained by the Top-N recommendation was recommended to the user.Finally,this paper conduct experiments in two open source datasets,and compared with several recommended algorithms.It is proved that time-aware hybrid point-ofinterest recommendation algorithm proposed in this paper has achieved good results.