:Individual driving destination prediction has important value in location-based services such as personalized services recommendation and intelligent transportation.However,most of the existing deep learning methods construct travel features on the basis of densely-sampled trajectory points,which generates data redundancy but only achieve limited information gain,and in turn introduces negative impacts on model training.Road intersection sequence can simplify the expression of driving trajectories that restricted on the roads,andreduce the model training cost.Meanwhile,transfer preference between intersections and current movement mode imply the spatiotemporal correlation between road intersections and destinations,and are capable to represent individual travel intention to some extent,but few researches have applied them to destination prediction. Methods
:The paper proposes a novel individual driving destination prediction method that takes Intersection Transfer Preference and Current Move Mode into account,named ITP-CMM.Firstly,it constructs input features in units of each intersection,and adopts Graph Attention Network (GAT) to learn the individual-level transfer coefficients between intersections in different time slots.Then,it uses Long Short-Term Memory (LSTM) to capture the long-term dependence of transfer preferences.Secondly,ITP-CMM constructs time cyclic encoding features and driving features to learn current move mode by using two-layered LSTM.Thirdly,intersection transfer preferences and current move mode are fused and weighted through feature intersection and attention mechanisms respectively.Then,the stacked residual networks are introduced to output prediction results.Based on the trajectory data of twelve private car drivers in Shenzhen City for the whole year of 2019,the paper has conducted a group of experiments to verify the feasibility of the proposed method ITP-CMM. Results and Conclusions
:We verify the effectiveness of the proposed method by carrying out ablation experiments and comparing the prediction accuracy and stability of ITP-CMM with four baselines,including Hidden Markov model (HMM),LSTM,Distant Neighboring Dependencies model (DND) and Attention-aware LSTM by considering Location Semantics and Location Importance (LSI-LSTM).Besides,we demonstrate the value of transfer preference in capturing the spatial correlation between intersections through visual analytics,and explore the impact of the number of transfer intersections on the prediction accuracy.