Research Of Federated Learning Methods For User Modeling

Massive personalized applications have gained popularity over the past decades,benefiting from the development of personal devices.Simultaneously,an unprecedented amount of data is being generated and stored on individual devices.The massive amount of personal data contain valuable user characteristics such as users’ interests,preferences and knowledge.Naturally,user modeling becomes a fundamental task,with the goal of analyzing behavioral information and inferring the unobservable characteristics in various intelligent technologies and applications.Generally speaking,centralized user modeling on collected data has raised concerns about the risk of data misuse and privacy leakage.For the purpose of protecting personal privacy and avoiding data leakage,federated user modeling method has come into favor,which expects to provide secure multi-client collaboration for user modeling through federated learning.Researchers also carry out a number of methods and applications for federated user modeling.However,there are still some limitations on the existing federated user modeling methods.To the best of our knowledge,existing federated learning methods that ignore the inconsistency among clients cannot be applied directly to practical user modeling scenarios,and moreover,they meet the following critical challenges:1)Statistical heterogeneity.The distributions of user data from different clients are not always independently identically distributed(IID),since differentiated usage scenarios and habits lead to unique clients with needful personalized information;2)Privacy heterogeneity.User data contains both public and private information,which have different levels of privacy,indicating that the different information shared and protected should be balanced;3)Model heterogeneity.The local user models trained with client records are heterogeneous,and thus require a flexible aggregation in the server;4)Quality heterogeneity.There are differences in the quality of user data in the clients,low-quality information from inconsistent clients poisons the reliability of user models and offsets the benefit from high-quality ones,meaning that the high-quality information should be augmented during the process.To address the challenges,this thesis develops a series of researches on federated user modeling for inconsistent clients from the following three aspects.With a primary goal of serving federated learning for user modeling in inconsistent clients,we first propose Hierarchical Personalized Federated Learning(HPFL)framework.Further,in order to augment high-quality information and generate high-quality user models,we expand HPFL to the Augmented-HPFL(AHPFL)framework by incorporating the augmented mechanisms,which filter out the low-quality information such as noise,sparse information and redundant information.In addition,considering the quality heterogeneity of education data in dynamic user modeling in education scenarios,we propose a federated deep knowledge tracking framework.The major work and contributions can be summarized as follow:1.We propose a federated learning framework for user modeling in inconsistent clients,HPFL.It follows the client-server architecture.More specifically,the clients train and deliver local user models via the hierarchical components containing hierarchical information from privacy heterogeneity to join collaboration in federated learning.Moreover,the clients update personalized user models with a fine-grained personalized update strategy for statistical heterogeneity.Correspondingly,the server flexibly aggregates hierarchical components from heterogeneous user models in the case of privacy and model heterogeneity with a differentiated component aggregation strategy.2.We propose an Augmented mechanism-based federated user modeling framework,Augmented-HPFL(AHPFL).It incorporates the augmented mechanisms,which filter out the low-quality information and augment the effect of high-quality information.Specially,we construct two implementations of AHPFL,i.e.,AHPFL-SVD and AHPFL-AE,where the augmented mechanisms follow SVD(Singular Value Decomposition)and AE(AutoEncoder),respectively.3.We propose a quality-oriented Federated Deep Knowledge Tracing(FDKT)framework.In this framework,each client takes charge of training a distributed dynamic user model,that is,deep knowledge tracing(DKT)model and evaluating data quality by leveraging its own local data,incorporating different education measurement theories.We construct two implementations based on FDKT,i.e.,FDKTCTT and FDKTIRTwhere the means of data quality evaluation follow Classical Test Theory and Item Response Theory,respectively.While a center server is responsible for aggregating models and updating the parameters for all the clients.4.Finally,we conduct extensive experiments on real-world datasets,which demonstrate the effectiveness of HPFL,AHPFL and FDKT frameworks.