A Trip Recommendation Method Based On Spatio-temporal Data Mining

In recent years,with the popularity of mobile terminal devices and the rapid development of social networks,location-based services have become an important part of people’s lives,and the resulting massive spatio-temporal data also provide us with unprecedented opportunities to explore and mine human mobility patterns.Opportunity.Among them,personalized trajectory recommendation as an important downstream task has become a current research hotspot.It is a very meaningful research to provide personalized itinerary recommendations according to user requests by mining the movement patterns and access preferences of people in trajectory data.Work.However,existing research still faces some key challenges: First,a large number of users have diverse travel plans due to different needs.Secondly,the complex relationship between points of interest(POI),including the geographical location and category characteristics of POI,and the semantic relationship of access time and context,how to integrate these complex heterogeneous relationships to effectively represent POI.In addition,problems such as insufficient representation and the curse of dimensionality caused by data sparsity.In response to the above challenges,we propose a method based on deep learning to conduct research on personalized trajectory recommendation.The main research contents are as follows:(1)This paper proposes a trajectory recommendation model(SelfTrip)based on self-supervised representation learning.SelfTrip mainly includes a POI representation learning module and a trajectory learning module based on contrastive learning.Among them,the POI representation learning module designs an enhanced POI probability transition graph and proposes a random walk strategy for trajectory data enhancement,while using a contrastive learning method to train the POI representation layer.In the trajectory learning module,we design four itinerary data augmentation methods,using contrastive learning to enhance the generalization ability of the model.Finally,a query representation-based recurrent neural network is used for trajectory generation,where the trip destination is enabled as a supervisory signal to constrain POI generation bias.Comparative experiments on 4 public datasets and our designed ablation experiments demonstrate the effectiveness of SelfTrip.(2)In this paper,we propose a dual-granularity trajectory recommendation model based on spatio-temporal graph fusion.The model uses graph information for POI representation,in which a heterogeneous spatio-temporal graph(ST-Graph)that integrates three different knowledges of time,space and semantics is constructed,and the graph convolutional neural network is used to extract spatio-temporal graph knowledge and used for POI embedding.Then,a GRU-based double-granularity recursive module is designed to capture the access preferences and transition rules of the crowd respectively,and then combine the popularity of POIs to determine the generation of POIs.Finally,for the cold start problem,we extract negative samples from the original data and the spatiotemporal map,and use the contrastive learning method to warm up the model.Experiments on 5 public datasets show that GraphTrip significantly outperforms baseline methods,and we employ a visualization approach to study the interpretability of graph knowledge.