Traffic Knowledge Graph Based Trajectory Destination Prediction

The task of predicting the final destination based on part of the trajectory of moving vehicle has great significance no matter in commercial advertisement recommendation,public security management or traffic dispatch.Since the road segments in traffic network are affected by the surrounding points of interest and functional areas,the result of destination prediction task depends on current trajectory status,time and other factors,and the sparsity characteristics of trajectory data may affect the accuracy of forecast.These factors have led to a very challenging task to learn travel intention of taxi from the known trajectory and accurately predict the destination.In order to solve above problems,this paper proposes a trajectory destination prediction model based on the traffic knowledge map.This model first builds a multi-layered traffic knowledge map from the bottom to top to model the complex traffic network and optimize related algorithms.The road network layer associates various road segment entities and interest point entities in the road network through positional relationships,Fusion of multi-source information in the road network to realize mutual perception between nodes in the road network.The trajectory layer can be used as a supplement to the road network layer to supplement the upstream and downstream direction and the probability of walking between the road segment entities for the relationship and the road segment entities in the road network layer,and establish the relationship between the road segment entities contained in the trajectory data between the entities.Relations,and learn the positioning of the road segment entity in the real trajectory.The functional layer uses the Mean-Shift algorithm to cluster the starting point and ending point of the trajectory to obtain the functional areas of each entity in the road network,and dig out the action machines behind the trajectory for subsequent destination prediction tasks.In order to embed the information in the multi-layer knowledge graph into the multi-dimensional space,this paper improves the Graph-Bert algorithm according to the characteristics of the multi-layer knowledge graph,and uses the weighted random walk algorithm to sum the correlation degree of the nodes in each layer.In this way,the knowledge graph is divided into multiple subgraphs and the characteristics of nodes in the subgraph and the whole graph are learned through Transformer.After obtaining the representation vector of the road segment entity,this paper uses the self-attention mechanism model to learn the node information and its timing information in the trajectory sequence.While retaining the information in the long trajectory sequence,the degree of contribution of different regions in the learning trajectory to the prediction result is used to make The forecast result has a more intuitive form.The time information and the functional area of the starting point are added to the model as metadata to improve the accuracy of model prediction.At the same time,in order to solve the problem of uneven label distribution caused by sparse trajectories,this paper also builds a binary search tree for the road network and divides the area based on the destination density.This article uses the Chengdu taxi trajectory data shared by Didi as a data set,and conducts ablation experiments on the modules in the model to prove the effectiveness of each module,and compares it with the Markov chain-based model and other neural network-based methods.The comparison and the best results were obtained.Finally,the robustness of the model in dealing with the sparsity of the trajectory was confirmed by changing the data distribution.