Federated Representation Learning With Data Heterogeneity For Human Mobility Prediction

With the development of science and technology and the popularity of information technology products,the application of smart wearable devices and location-based smart services has brought a new paradigm for intelligent human mobility prediction(HMP)and has a wide application in smart healthcare,intelligent cities and smart transportation.Existing methods are often based on deep learning models that can effectively exploit information from users to achieve more accurate predictions for better services.The traditional model training methods upload the user’s data to the server for training in a unified manner.however,with the implementation of data protection and other measures.strict data regulations have made people more and more aware of privacy preservation.The proposed Federated Learning Paradigm enables the protection of user privacy to some extent by keeping the data local for training and collaboratively training HMP models without sharing the highly sensitive location data with others.However,there are still two very important and urgent challenges of data heterogeneity in the real users’ trajectory mobility prediction problems.One is the heterogeneity of users’ mobility patterns.Due to the influence of demographic and other objective factors,the trajectory patterns of different individual users are highly heterogeneous and the distribution of the raw data is non-independently homogeneous;secondly,the data available for training is very scarce and there is the problem of insufficient labels or limited trajectory data.Therefore,this paper proposes an end-to-end federated representation learning framework for users’mobility prediction,called FR-HMP,to address all the above challenges.Firstly,this paper designs user-local trajectory prediction models that can dynamically capture temporal as well as positional information of historical mobile trajectories,and encoder-decoder structures that can enhance the representation of trajectory vectors.A dual-server module,consisting of a parameter server and a third-party server,is designed in this paper to address the potential privacy leakage issues associated with the simultaneous upload of latent representation information and model parameter information.In particular,the server-side module is a two-stage process in which the clustering module on the parameter server can aggregate similar clients together and execute a federated average algorithm within clusters to address the problem of heterogeneous mobility patterns of users:while the representation learning module adaptively learns sinilar graph structures among users through a graph learning layer on the third-party server,and later uses the graph convolution layer and similar graph structures to learn enhanced representation of each client’s trajectory information to tackle the problem of data scarcity of users and improve the model training effect for data-scarce users.Through split learning,the training of model parameters without raw data on the server side is realized in order to achieve optimal privacy protection.Finally,extensive experiments were conducted using two different real-world publicly available human mobility trajectory datasets,and the results showed that FR-HMP achieves optimal results compared to existing state-of-the-art methods.In particular,FR-HMP achieves superior results when the number of predicted location classes is large.In addition,ablation experiments are conducted to verify the necessity and effectiveness of the modules of the model framework,and the clustering results and convergence process of the model are demonstrated.Furthermore,the hyper-parameter analysis of the Topk value of the graph learning layer,the size of the representation vector dimension,and the number of server-side representation learning training rounds provides criteria for the selection of the initial parameters of the model,and the communication and computational costs set the foundation for the next work.Finally,the paper gives constructive ideas for the next work.