Human Mobility And Social Knowledge Discovery Via Deep Trajectory Learning

Recently,the rapid development in mobile internet and the wide use of global position systems generate a plethora of location-based applications,which make information interaction among humans more frequent and diverse.And these applications provide unprecedented opportunities for the user to share personal experience and interesting locations,make friends with people who have similar interests,join a ride-sharing trip,etc.For instance,location-based applications,e.g.,We Chat,Twitter,and Weibo,usually collect large amounts of footprints(check-ins)left by users,social friends and etc.And the collected data can further yield valuable information such as personal trajectory and individual social relationships.Although such information is from the virtual world,i.e.the internet,it is the objective reflection of human activities from the real world.Currently,learning human trajectories has attracted the attention of researchers and practitioners,whereafter proposing various excellent models based on trajectory learning to capture human spatio-temporal information.However,traditional approaches are still confronted with three challenges during trajectory learning.First,the diversity of human mobility shows the difference in diverse moving patterns,location favorites,time preferences,etc.Second,the sparsity of data such as the sparsity in check-in data brings inadequate representation of underlying features and difficulty in obtaining personal preference.Besides,the semantic complexity of trajectory should be addressed in trajectory modeling.Lately,deep learning technique with powerful ability of generalization and feature extraction has achieved huge success in many areas such as natural language processing and image processing.Consequently,it provides us a new perspective to address the limitations in trajectory learning.For example,applying the recurrent neural network to capture the long dependency in trajectories.In summary,this dissertation aims to discover the knowledge of human mobility and social relationships via deep trajectory learning,along with addressing the aforementioned challenges.In particular,introducing several deep learning techniques to understand human trajectories and tackling four significant but challenging tasks,i.e.,human mobility identification,next POI prediction,trip recommendation,and social circle inference.The main contents of this dissertation are presented as follows:(1)For human mobility identification,a novel trajectory classification problem is formulated that is linking trajectories to users who generate them,i.e.,Trajectory-User Linking(TUL).Since TUL is a typical trajectory classification problem,a Recurrent Neural Networks(RNN)based semi-supervised learning model,called TULER(TUL via Embedding and RNN)is proposed,which exploits the spatio-temporal data to capture the underlying semantics of user mobility patterns.And TULER embeds each check-in into a low-dimensional space based on the observation that the frequency of location checkin follows a power-law distribution.Subsequently,TULER leverages check-in embeddings and the recurrent neural network to achieve trajectory information capture and classification training in a semi-supervised manner.In the end,experiments conducted on real-world datasets demonstrate that TULER achieves better accuracy than the existing methods.(2)For next POI prediction,A variational attention-based next POI prediction model(VANext)is proposed to address the limits in POI prediction including the sparsity of recent mobility,the density of historical mobility,and the complexity in data.VANext is a latent variable model for inferring user's next footprint,with historical mobility attention.The variational encoding in VANext captures the latent features of recent mobility,followed by searching the corresponding historical trajectories for periodical patterns.A trajectory convolutional network is then used to learn historical mobility,significantly improving the efficiency over often used recurrent networks.Finally,VANext predicts the next POI by exploiting the periodicity of historical mobility patterns,combined with recent check-in preference.Experiments conducted on real-world datasets demonstrate that proposed methods outperform the state-of-the-art human mobility prediction models.(3)For trip recommendation,a novel model,namely Deep Trip,is proposed to address the trip recommendation problem under a framework of encoder-decoder,jointly using an adversarial neural network to train a code space for improving the recommendation of preferred routes.Deep Trip consists of: a Trip Encoder to embed the contextual route into a latent variable with a recurrent neural network(RNN);and a Trip Decoder to reconstruct this route conditioned on an optimized latent space.Simultaneously,defining an Adversarial Net composed of a generator and critic,which generates a representation for a given query and uses a critic to distinguish the trip representation generated from Trip Encoder and query representation obtained from Adversarial Net.Deep Trip enables regularizing of the latent space and generalizing users' complex check-in preference.The theoretical interpretation of Deep Trip and extensive experimental evaluations on two realworld datasets demonstrate the effectiveness and robustness of Deep Trip in comparison to the state-of-the-art baselines.(4)For social circle inference,an essential definition of trajectory-based social inference(TSCI)problem is given.TSCI aims at inferring user social circles(mainly social friendship)based on motion trajectories and without any explicit social networked information,where this dissertation will formulate it as a novel multi-label classification problem and then develop the Recurrent Neural Network-based framework called Deep TSCI to use human mobility patterns for inferring corresponding social circles.And three variants of Deep TSCI are presented to learn the latent representations of trajectories,based on(1)Bi-directional Long Short-Term Memory(LSTM);(2)Autoencoder;and(3)Variational autoencoder.Experiments conducted on real-world datasets demonstrate that our proposed methods perform well and achieve significant improvement in terms of macro-R,macro-F1 and accuracy when compared to baselines.