Key Technology Research On Point Of Interest Based On Review Information To Recommend

With the widespread use of smart phones and the popularity of mobile Social platforms in recent years,location-based Social Networks(LBSN)business has been developing rapidly,including Point-of-Interest(POI)recommendation.The POI recommendation uses a usergenerated check-in history to predict the POIs that the user is interested in,and to personalize the POIs that the user will visit at the next moment in a large set of POI candidates.This makes the user personalized travel experience has been better improved.The main research goal of this paper is to better model the characteristics of POI and users' interests based on the comment information,and to provide users with a more accurate and personalized recommendation list of POI.Due to the sparse check-in data,how to better integrate review information becomes a big challenge for POI recommendation.Among the existing methods with the best effect,although the goal of integrating comment information in POI recommendation is achieved through independent modeling of POI sequence features and comment information sequence features,the process of interaction between POI information and comment information is lack,and POI embedding vector lacks dynamic semantic representation in different contexts.As a result,the extraction of POI features and user interest features is not sufficient.This paper proposes a two-stage model architecture,which is divided into pre-training and fine-tuning,and a neural network model based on attention mechanism to predict the POI that users may visit at the next moment.Firstly,in the pre-training stage,a context-based location embedding model is proposed,which adaptively generates a context-based dynamic expression embedding vector for the target location.At the same time,the target vector is trained by combining the comment word context with POI to obtain the POI vector that integrates the comment information.It not only enables embedded vectors to express dynamic semantic features in different contexts,but also enables POI vectors to integrate certain semantic information.Secondly,in the fine-tuning stage,a novel multi-factor attention mechanism module is proposed,which captures sequence correlation within the same mode,semantic correlation between different modes,and semantic-based second order correlation within the same mode.It can not only effectively realize the synchronous interaction of information between modes,but also realize the adaptive feature selection about the correlation between sequence information and semantic information within modes,which makes the information feature extraction more sufficient.Then,the aggregation attention mechanism module is used to aggregate the information of different factors in the mode,so that the model can adaptively aggregate multi-factor information in different contexts.In this paper,we analyze the weight matrix of multiple sign-on tracks in the attentional mechanism through experiments on a sign-on dataset containing comment information in the real world,and further prove the role of multi-factor attentional mechanism module in the prediction process.Finally,the prediction performance of the proposed model is compared with that of other benchmark models.The final experimental results show that the prediction accuracy of the proposed model is better than that of the benchmark model,which proves that the proposed model can effectively integrate comment information and better model POI recommendation problems.