Research And Application Of Deep Representation Learning With Label Noise Based On Mobile Trajectory Prediction

Mobile trajectory prediction plays an important role in many cityrelated fields such as urban planning,traffic engineering,business decision-making,etc.Thanks to the fast development of mobile Internet and big data technology,it is now possible to capture the whole city-wide trajectory data.Besides,the application of deep learning technology has also improved the ability to model human mobility.However,existing researches mainly collect data from volunteers with specific equipment,or mobile application check-in logs,which often cover thousands to hundreds of thousands people and lack of representativeness.In addition,due to the complexity and lack of authentic credible label references,the noise carried in trajectory data is usually hard to clean.Artificially setting fixed metric thresholds or cleaning rules may even lose valid data,which further reduces the performance of mobility prediction model.To solve above problems,this thesis mainly focuses on improving the performance of mobile trajectory prediction when dealing with large-scale noisy trajectories in real world.This thesis decomposes it into three subproblems:1)The conventional trajectory prediction model,which takes the user’s former trajectory data as input,and predicts the user’s next location;2)How to identify noisy labels,or,how to quantify the accuracy of a trajectory based on its labels;3)How to utilize the quantitative results obtained in 2)based on 1)to improve the prediction performance in the overall scenario.The main work of this thesis is as follows:Firstly,based on the deep learning model,this thesis implements a complete process from distributed log data processing to mobile trajectory prediction.A prediction model with combination of Bidirectional Gated Recurrent Unit and attention mechanism is also proposed,which has been validated on large-scale noisy trajectories in real world,to improve the accuracy of prediction.At the same time,the label noise problem of trajectory data is analyzed,and its influence on the prediction model is verified.Secondly,for the identification of label noise,this thesis proposes a calibration model of user trajectory data with its unsupervised training method based on deep representation learning and verifies the rationality of its numerical evaluation of label noise.Finally,by using the idea of Instance Weighting to calibrate the neural network parameters,this thesis combines the above models to propose a complete mobile prediction method with a calibration network,which effectively improves the performance of mobile trajectory prediction based on large-scale noisy trajectories in real world.