User Mobile Pattern Mining Based On Meta Learning And Variational Generation Network

In the era of the Internet of things(IoT),an enormous amount of mobile devices collect a large volume of personal mobility data,which allows the learning of explicit and implicit patterns of human beings.Mining human mobility data can not only benefit many business applications,but also assist various smart city services.though having shown promising direction and encouraging results,existing methods require large number of labeled data for training an effective task-specific model,and thus are not applicable to other scenarios where only a few samples are observed or the data types have not been encountered during training.To overcome the above challenges,firstly,we present a Semi-supervised Trajectory-User Linking model with Interpretable representation and Gaussian mixture prior(STULIG)– a novel deep probabilistic framework for jointly learning disentangled representation of user trajectories in a semi-supervised manner and tackling the location recovery problem;secondly,we present a novel human mobility learning methodology,called MetaMove,which is the first meta-learning based model generalizing the mobility prediction and classification problems in a unified framework.MetaMove addresses the task generalization problem by training the model over a variety of learning tasks,sampled from different users,and the model is optimized on a distribution of tasks the main contributions of our model are as follows:(1)We propose a model based semi-supervised learning and Gaussian mixture prior which uses the spatial and temporal information of users and unlabeled data to improve the performance on TUL.(2)We propose a general framework MetaMove based on meta learning and deep generation network,which can be easily combined with the existing human mobility learning methods.Our model can use unlabeled data in meta training and adaptive level to alleviate the problem of data sparsity,while forcing the model to be less sensitive to negative samples.(3)We conduct extensive experiments on two practical applications,and the results demonstrate the effectiveness and efficiency of our two methods,achieving significant improvement over the state-of-the-art approaches.