Personalized Trajectory Learning At Designated Way-points

With the popularity of GPS,trajectory data that implies user preferences have been widely used in intelligent transportation,cities,logistics and other fields.Compared with traditional optimal routing,trajectory based routing can better capture multidimensional personality preferences and has become a research hotspot.Route planning for services such as express delivery,food delivery,and ride-sharing involves personalized routing planning that requires a given origin-and-destination pair to pass through a series of designated way-points,an issue that has not received much attention.The research on this problem can not only optimize the above services,but also help to save energy and reduce carbon,which has theoretical significance and practical value.Personalized route planning with designated way-points faces difficulties in the following two aspects: 1)Users’ local trajectory selection at different times is different due to their personal habits.How to give consideration to individual preference characteristics in trajectory learning? 2)The way-point is a regional concept with uncertainty.How to determine the trajectory based on the fuzzy way-point region?In fact,the sequence composed of origin-and-destination pair and way-points can be called rough trajectory,while the real trajectory can be called thin trajectory.The two are correlated.Therefore,trajectory learning at designated way-points can be regarded as a machine translation problem,that is,translating rough trajectory into fine trajectory.However,the existing trajectory representation framework based on Seq2 Seq does not consider the personalized requirements and fuzzy way-point region characteristics.In order to realize personalized trajectory learning at designated way-points,this paper proposes trajectory search model and trajectory generation model,with specific contributions as follows:(ⅰ)A Personalized Selective Route(PSR)model is proposed to realize Personalized trajectory search at designated way-points.The dynamic personalized preferences of users are learned by integrating multi-source information such as user and time in trajectory representation.A multi-head self-attention mechanism is designed to capture the features of longer trajectories,and a recommendation module is further designed to realize the fine trajectory search similar to the coarse trajectory in the designated way-point region.Experiments on real traffic trajectory data verify the PSR model’s recommendation accuracy and robustness.Compared with baseline LCSS model,the accuracy of PSR improves 32.57% and 63.87% on top-1 and Top-3recommended trajectory routes,respectively.(ⅱ)A short-term Goal Learning for Trajectory Refinement(STGTR-TRAj)model is proposed to estimate short-term goals based on way-point and complete the generation of fine trajectories segment by segment.The endpoints-CVAE are designed to capture the local context features,learn the potential Gaussian distribution of short-term target around the local fine trajectory,and estimate the user’s preferred short-term target set in the way-point area.The similarity attention network was designed to capture the global context features and describe the spatial shape similarity between coarse trajectory and fine trajectory.Furthermore,a segment-by-segment fine trajectory estimation module is designed to capture the influence of different short-term target estimates on trajectory refinement.Experiments on the real trajectory data set show that STGTR-traj improves the Jaccard similarity,minimum correction distance and maximum cumulative distance by 7.54%,7.41% and 16.52%,respectively,compared with the optimal baseline.The above two models are suitable for different situations: the former is suitable for situations with rich historical trajectories,while the latter can generate personalized trajectories that meet the requirements of designated way-points without directly corresponding historical trajectories.33 figures,17 tables,and 59 reference articles are contained in the dissertation.