| The proliferation of mobile terminals,leads to the generation of large-scale trajectories with an unprecedented rate.As the fundamental and an indispensable component of trajectory data mining,trajectory representation learning has attracted more and more researchers’attention in recent years.However,trajectories in different scenarios have different attributes(e.g.trajectories in road networks are constrained by the topological structure of road networks),and existing TRL methods have poor performance in modeling long trajectories.Hence,the universality of the representations of trajectories are challenged.In this context,based on the neural network method,this paper studies the trajectory representation learning in road network scenarios and non-road network scenarios and its performance in similarity query tasks.The main contents are as follows:(1)In order to effectively characterize trajectories with different lengths on road networks,the model namely TRMS is proposed in this paper.Firstly,extends the traditional algorithm Deep Walk with a probability-aware walk which considers the transition probability from a road segment to each of its connected road segments,fully capturing the topological structure of road networks,and then utilizes the masked seq2seq learning framework and self-attention mechanism in a unified manner to capture the long-term semantic features of trajectories.Based on the next position prediction and destination prediction tasks,experiments conducted on the real-world datasets demonstrate that TRMS outperforms the state-of-the-art methods in embedding short,medium,and long trajectories.(2)In order to learn robust representations for trajectories with different lengths in nonroad network scenarios,the model DGTM based on dual attention mechanism is proposed in this paper.Firstly,the model combines the Deep-walk algorithm and the graph attention network,and simultaneously considers the context information and neighborhood information of grid cell,to effectively capture node features.In trajectory representation learning,DGTM dynamically constructs grid sequence subgraphs of trajectories based on each batch data.On each subgraph,local features are deeply mined based on graph attention mechanism and long-term semantic features of trajectory are captured based on self-attention mechanism.Based on the next position prediction and destination prediction tasks,experiments conducted on the real-world datasets demonstrate that DGTM outperforms the state-of-the-art methods in embedding short,medium,and long trajectories.(3)This paper further verifies the effectiveness of TRMS and DGTM based on the trajectory similarity query task.In order to shorten the query time,this paper firstly clustering the trajectories in the database according to the distribution density based on DBSCAN algorithm to reduce the candidate set size of the query trajectories.Then,the feature representations of trajectories are obtained according to above trajectory representation learning model,and the similarity between trajectories is represented by calculating the distance between vectors.The Mean Rank、HR@10、HR@50 and R5@20 of the Ground-Truth corresponding to the query trajectory in the database are taken as evaluation indicators.Finally,experiments conducted on the real-world datasets demonstrate the superiority of DGTM and TRMS in similarity query task.To summarize,this paper proposes two trajectory representation learning models based on neural network methods to effectively characterize trajectories with different lengths in road network scenarios and non-road network scenarios,which is of great significance for enriching relevant researches on trajectory mining. |
