The Research For Service Composition And Recommendation In Pervasive Environments

Pervasive service extends the concept of traditional web-based service by the abstraction of the resources on pervasive devices as services.The emergence of various wireless network technology gives users the superior computing capability over single devices in a distributed pervasive environment.However,the discovery methods for single services are insufficient in the satisfaction of users with the ever increasing requirement for complex services.Additionally,the lack of the understanding for users' preferences leads to the difficulties on the satisfaction of users' personalized service requirement,which further decreases the overall experiences.Therefore,service composition and recommendation related technologies are introduced to solve the deficiencies.Based on service discovery related technology,pervasive service composition provides users with complex computing services through the coordination of the distributed sub-services.Currently,the following problems should be solved during the optimization of the network topology for composite services.First,there exists multiple sub-services for the implementation of one sub-functions,which results in the exponentially increasing number of the composition solutions.Second,the universal existence of parallel branches in a functional graph hinders the direct application of conventional optimization methods for service composition.Third,there exists none centralized node that is responsible for the cache of the services on other nodes,which determines a distributed manner for the optimization.Fourth,the mobility of pervasive nodes requires both the efficiency and the reliability of the service composition to be considered.This thesis first solve the problem for path-structured service composition,which models the functional graph and the solution as a directed path.Based on distributed message forwarding,ForeNoRepSearch and ForeBackSearch algorithms are proposed.ForeNoRepSearch incorporates path filtering to discard non-optimal partial solution in early phases.On the basis of path filtering,ForeBackSearch introduces path combination,which start the sequential search for the sub-services with the first and the last sub-function of the functional path.The experiments on a simulation platform for distributed service environments show the outperformance of ForeNoRepSearch and ForeBackSearch over conventional methods on the searching efficiency and the solution quality.The results also validate the superior efficiency of ForeBackSearch over ForeNoRepSearch for the introduction of path combination.For the parallel branches in the functional graph of service composition,a 3-staged optimization approach is proposed.This approach first identifies all the path-structured sub-graphs in a functional graph through a depth-first traversal on the graph.The sub-graphs are forwarded among the pervasive nodes to search for the corresponding sub-service compositions,with path filtering and relay path optimization introduced,and these compositions are gathered onto the service requesting node.The third stage is abstracted as a Constraint Optimization problem,and a Branch-and-Bound method is proposed to compute the optimal solution.For the frequent pervasive node mobility,the reliability of a service composition is evaluated by the residual lifetime of the links in the composition,and Pareto optimization is introduced to take both the delay and reliability metric of the composition into account for the application of the 3-staged approach.The thesis validate the performance of the 3-staged approach for various kinds of functional graph topologies under static and dynamic experimental environments.The results show that this approach solve the problem of the service composition with parallel branches effectively.Pervasive service recommendation predict users' personalized preference for services by mining the historical interaction data between users and services.Conventional collaborative filtering-based methods suffer from data sparsity,therefore gives low prediction accuracy.Matrix Factorization(MF),as an effective dimension reduction method,relieves the sparsity problem by modeling the factorized user and service latent factors.However,the effectiveness of MF is limieted by the one-class setting of the data and the non-informative priors on latent factors.Therefore,this thesis introduces kernel functions to construct user-user and service-service neighboring relations by the combination of user-service and service-category data.The neighboring relations are incorporated into Probabilistic Matrix Factorization(PMF),which forms Kernel Constrained PMF(KCPMF)and Kernel Regularized PMF(KRPMF)respectively.KCPMF decomposes the latent factors into two components,personalized offset and neighboring effects,while KRPMF regularizes the latent factors with the linear combination of their neighbors.The experiments on a real-world user-app installation dataset show that KCPMF and KRPMF relieve the negative effects of one-class setting on the prediction,compared with conventional CF and MF baselines.To further mine the user-service personalized preferences,this thesis maps the duration in user-service interaction information to pseudo ratings,to capture the users' preference degrees for different services.Based on the pseudo ratings,a Kenelized Non-negative Matrix Factorization(KNMF)is proposed,which places non-negative constraints on the latent factors and adopts a Gaussian Process(GP)model over the latent factors.The parameters of the GP is obtained by the application of Radius Basis Function on the pseudo ratings and service-category data,to present fine-grained user-user and service-service similarities.For the non-negative constraints,KKT conditions are introduced to form the Multiplicative Updating Rules-based gradient descent.An auxiliary function is introduced to prove the convergence of MUR on the optimization of KNMF.The experiments on the user-app dataset show that KNMF improve the recommendation accuracy and learning efficiency evidently,compared with conventional PMF and NMF models.