Research On Point Of Interest Recommendation Method Based On User Intention And Sequential Preference Features

The development of Internet and mobile communication techniques provides massive information and resources to people’s life.Recommendation systems are playing an increasingly important role in helping users filter irrelevant information and assisting decision-making.Point of Interest(POI)recommendation connects the real life and the digital world.By analyzing user’s behaviors in digital world,understanding and mining user’s preferences,POI recommendation recommends location resources that meet user’s preferences in real life.It has become an important service in location-based social networks.Scholars have made some explorations on POI recommendation and accumulated some research results.However,with the background of massive data and complex scenes,user’s preferences show the characteristics of diversification and personalization,which leads to some new challenges for POI recommendation:How to develop new POI recommendation method that adapts to the complex scenes by integrating the internal and external influencing factors of user’s behavior.How to explore the influence of internal factors such as user’s intentions on user’s behavior.How to learn personalized sequential features in user’s behavior to grasp future trends,etc.To address these challenges,this paper carried out the following work:(1)A novel POI recommendation method that characterizes user intentions using POI category and considers spatio-temporal preferences is proposed.Based on existing work,we use POI category to describe user’s intention,and establish an intention preference learning method.Simultaneously,we propose a sequential preference learning method based on LSTM,to predict the check-in probability in time dimension.A personalized location preference modeling method based on kernel density estimation is proposed,to predict the check-in probability in spatial dimension.By integrating the check-in probability in intention,sequential and spatial dimensions,we obtain the candidate POIs and organically integrate internal and external factors in POI recommendation.Experiments on real check-in datasets verify the superiority of our method compared to other methods on POI recommendation tasks.Simultaneously,we explore the influence of different factors on recommendation performance,and evaluate its importance.(2)A hypersphere model for describing user’s preferences and a novel POI recommendation method are proposed.To explore user’s temporal preferences in their behavior,we build a check-in graph using historical check-in records.Then,we adopt DeepWalk algorithm to learn the presentations of POIs,building a hypersphere preference model by generating user interest center and interest radius,mapping user preferences to high-dimensional feature space.Meanwhile,we established a sequential preference learning method based on a stacked neural network,using Bi-LSTM,self-attention network and Memory network.Using the hypersphere preference model,we propose a data-adapted conversion function to convert the distance of candidate POI representations and the interest center to check-in probability.It is used to revise the ranking score of candidate POIs to improve the recommendation performance.Experiments on real check-in datasets indicate that our method outperforms the comparison methods on multiple evaluation metrics.In addition,we discuss the importance of different components and explore the impact of important parameters on recommendation performance in our experiments.(3)A intention learning and POI recommendation method based on temporalhierarchical intention features is proposed.Considering the hierarchical POI categories,we propose a user intention representation method by building a hierarchical intention tree.We adopt LSTM and attention network to build an intention learning method,to learn user intentions from both temporal and hierarchical perspectives.In order to mine the sequential preference features of user’s behavior,we build a sequential preference learning method based on Memory network,and propose a novel POI recommendation method by learning intention and sequential preference features.Simultaneously,we build an intention prediction module,which adopts a multi-task learning strategy to train the model to predict future intention and check-in behavior,to improve POI recommendation performance by predicting future intention.The experimental results indicate that the proposed method not only improves the performance of POI recommendation,but also achieves accurate prediction of check-in intention.