Human Trajectory Analysis Based On Self-supervised Learning

With the widespread of smartphones and location based services(LBS)software,a growing number of people begin to use these applications to check in the places they have visited,thus accumulating a large number of mobile trajectory data and providing the unprecedented opportunity to learn user mobile patterns from the trajectory data,which will be conducive to a series of business and management applications,such as location recommendation,anomaly trajectory detection,crime identification,epidemic disease tracking,etc.Aiming at the sparsity and noise of user trajectory data and the existence of implicit feedback and weak supervision in existing models,this thesis proposes a framework based on Self-supervised Mobility Learning(SML).The framework can be applied to user Location Prediction(SML-LP)and Trajectory User Classification(SML-TUL).First,the SML framework mainly uses Pretext Tasks to mine the supervision information of the data itself from large-scale unlabeled data.By constructing supervision signals to train the model,it can learn useful representations for downstream tasks.Second,SML is designed for modeling sparse user trajectory data,which can enhance the generalization ability of the model by leveraging rich spatio-temporal contextual information and augmented data to improve trajectory representation.Third,SML provides a principled approach to describe intrinsic motion correlations while addressing the implicit feedback and weak supervision issues in the existing models.Fourth,SML introduces a comparative instance discrimination method in the training of spatio-temporal data,which explicitly distinguishes the real user check-in points from the negative sample check-in points that are easily mispredicted.Aiming at the poor performance of existing models and the existence of NegativePositive-Coupling effects(NPC)in the contrastive loss function,this thesis proposes a framework based on Decoupled Contrastive Mobility Learning(DCML),the framework is applied to the Location Prediction task(DCML-LP).First,DCML-LP solves the NPC effect in the positive and negative samples of the trajectory by decoupling the objective function of contrastive learning(DCL).This decoupling operation significantly improves the efficiency of self-supervised mobile learning.Second,DCML-LP removes the NPC multiplier,so that the model does not need to rely on a large batch size to achieve its competitive performance,therefore,DCML-LP alleviates the requirement for computing.This thesis evaluates these models on multiple real-world datasets and compares the proposed SML and DCML frameworks with previous benchmark methods.The experimental results show that the performance of these two methods in this thesis is superior and can improve the accuracy of location based services.