Research And Implementation Of Traffic Structure Mining Model Based On Multi-Task Learning

As an important reference for urban traffic dispatching and individual travel,the compositi.on of urban traffic structure is an important prerequisite for realizing intelligent transportation system better at present and in the future.On the basis of existing traffic structure mining models and research theories,this paper focuses on their shortcomings:(1)taking the whole city as the research object,emphasizing the impact of urban traffic policy;(2)traditional machine learning method relies on a large number of traj ectories,which can not make full use of information as well,resulting in information waste or irrelevant information redundancy;(3)mining traffic structure for a specific space is often neglected because of its small space range and small number of trajectories,etc.Therefore,mining models suitable for different scenarios with different structures are proposed in this paper,based on the idea of multi-task learning.The main research contents are as follows.1.The basic theories and key technologies related to traffic structure analysis are summarized respectively.Then,existing research models are explained and summarized,the shortcomings of which are analyzed.Finally,the research ideas and concrete framework of traffic structure mining based on mesoscopic level are put forward.2.Considering the situation that the partition of urban space is coarse,the spatial range is large,and the amount of the trajectory is relatively sufficient,or when a specific area is not designated,the purpose of which is to reveal the influence of user-based feature attributes on traffic structure distribution,etc.This topic combines traditional machine learning with multi-task learning,and constructs a traffic structure mining model for coarse-grained partition of urban by utilizing the prior relationship of graph structure between features,which makes the full use of data,avoids unnecessary information waste and interference of redundant information.Meanwhile,comparing it with the classical multivariate logistic regression model.3.In order to mine and analyze the traffic structure composition in a small specific spatial scope,or considering situation that small amount of trajectory and fine-grained partition of the city,this topic takes migration learning and multi-task learning as the guiding ideology,a two-dimensional traffic structure mining model based on shared matrix structure for simple space environment and a three-dimensional traffic structure mining model for complex space environment are proposed respectively,both of which are based on the reality of this subject,utilize the geographic location correlation among multiple single areas,and the correlation among transportation modes,etc.what is important is providing a new solution for the analysis and research of traffic structure composition.