Research And Implementation Of Vehicle Destination Prediction Based On Embedding Learning

In recent years,the automobile industry has entered the era of experience economy,and the demand for intelligent vehicles is increasing.The human-computer interaction of automobile has changed greatly,and in-car systems have become a key factor affecting user experience.Pre-trip destination prediction can predict destinations for users before they leave,which is convenient for vehicle-mounted systems to provide smarter services such as location recommendation,route planning,and feedback on road conditions.At the same time,pre-trip destination prediction is of great significance for the analysis of travel distribution and the auxiliary of urban traffic system planning,so it has gained world-wide attention from researchers and industries.Current pre-trip destination prediction researches focus on using the mapping relationship between locations,ignoring users’ personalized travel characteristics,which is crucial to the prediction task.In addition,these methods often predict based on the statistical probabilities of users’ historical destinations,which is difficult to predict destinations that users rarely or never visited,so the prediction accuracy is limited.This thesis aims at solving the above problems to improve the accuracy of pre-trip destination prediction.Solving these problems has the following challenges:(1)Mining complex user-location mapping relationships from sparse data,and effective modeling of users’ personalized travel patterns;(2)Extracting the key information in the trip triplet composed of user,time and location;(3)Effective mining of users’ deep travel information in the vehicle travel network.For these difficulties,this thesis proposes vehicle destination prediction models based on embedding learning,which represent user travel characteristic and geographic location characteristic in the form of embedding,and generates prediction results based on node embeddings.The main contributions of this thesis are as follows:(1)This thesis introduces the research background,research significance,and current status of domestic and foreign research on the subject of vehicle travel destination prediction.By summarizing the advantages and disadvantages of the existing methods and analyzing the time and space laws of users’ travel in the dataset,the design ideas of the model in this thesis are proposed.(2)This thesis proposes a destination prediction model based on embedding combination learning,which realizes the modeling of the characteristics of users,time and locations through the embedding learning of each tuple in the trip triplet(user,departure time,departure place)to solve the first difficulty.This thesis proposes an embedding combination mechanism based on multi-layer perceptron,which extracts the key information combination from each tuple embedding,and forms a combined embedding representation of trip triple to solve the second difficulty,and designs destination probability ranking optimization method to update node embeddings and other model parameters.Finally,the destination prediction results are generated based on the matching probabilities of trip triplet embedding and embeddings of candidate destinations,and the effectiveness of the proposed model is verified through experimental comparison.(3)This thesis proposes a destination prediction model based on embedding propagation learning,which designs an embedding propagation mechanism based on the idea of graph neural network to solve the third difficulty.The model propagates the neighborhood information to the learning process of node embedding,and mines the deep interactive information between nodes in the vehicle travel network.This thesis proposes a node correlation ranking optimization method to update the model parameters.Finally,the embeddings of trip triples are input into prediction models to generate the prediction results,and the effectiveness of the proposed model is verified through experimental comparison.(4)According to the proposed models,we research and design a corresponding vehicle travel destination prediction system,which includes functions such as vehicle travel destination prediction and POI recommendation,to achieve destination prediction before users travel and recommend places of interest to them.