Research On Temporal-aware QoS Prediction Method Based On Non-negative Latent Factorization Of Tensors

Quality of Service(QoS)is a crucial evaluation criterion for measuring web-service performance in web-services.When measuring a web-service is suitable or not for a user to invoke a specified service at a specified time point,it is vital to conduct data mining and data analysis on historical QoS data with rich information and predict temporal-aware QoS missing data based on these hidden information.Recently,a Non-negative Latent Factorization of Tensors(NLFT)–based algorithm is the state-of-the-art one for capturing hidden temporal-aware QoS information.It employs tensor modeling to represent the user-service-time relationship and achieves a Single Latent Factor-dependent,Non-negative and Multiplicative Update on Tensors(SLF-NMUT)learning rule by manipulating learning rates of a algorithm to cancel negative terms during the itearition process.An SLF-NMUT learning rule non-negatively constraints decision parameters of a learning algorithm,making temporal-aware QoS missing value prediction have practical significance of non-negativety.However,this learning rule can easily lead to uncontrollable learning rates of an algorithm during the optimization process,making a algorithm easy to oscillate in the iterative process,suffering from the trouble of low-tail convergence and precision loss.Hence,this paper based on a NLFT algorithm proposes a series of optimization algorithms from the points of improving prediction accuracy and accelerating converfence.The main contributions of this paper include the following:(1)We propose exponential rescaling,linear rescaling,and multi-rescaling SLF-NMUT learning rules by adopting learning depth rescaling strategy.These learning rules can effectively control learning depths of a learning algorithm during the iterative process,making models to optimize objective more precisely.A NLFT-based algorithm adopts these learning rules for iterative training and obtains the three NLFT models with exponential rescaling,linear rescaling and multi-rescaling learning depth.Empirical studies on two industrial datasets demonstrate that the proposed algorihtms effectively control its learning depth,capture temporal information of temporal-aware QoS data,accurately predict missing QoS data,and improve the prediction accuracy of models.(2)Based on the classical momentum method and the Nesterov accelerate gradient(NAG)method,this paper proposes the generalized momentum method and the generalized NAG method.The standard momentum method and the NAG method introduce the velocity vector of the previous iteration to accelerate the convergence of a learning algorithm,and they are explicitly dependent on the gradient terms of a learning algorithm.However,a NLFT algorithm adopting the SLF-NMUT learning rule that manipulates the learning rates of the a learning algorithm implicitly depends on the gradient terms of a learning algorithm,which does not match the standard momentum method and the standard NAG method that explicitly depend on the gradient terms of a learning algorithm.Hence,this paper proposes a generalized momentum method and a generalized NAG method which implicitly rely on gradient terms,and achives a Single Latent Factor-dependent,Non-negative and Multiplicative and Momentum-incorporated Update on Tensors(SLF-NM2UT)learning rule and a Single Latent Factor-dependent,Non-negative and Multiplicative and Nesterov-incorporated Update(SLF-N~2MUT)learning rule by combining these two methods with SLF-NMUT learning rules.An NLFT-based algorithm trains with SLF-NM~2UT and SLF-N~2MUT learning rules,two fast convergence NLFT models are proposed:Momentum-incorporated Non-negative Latent Factorization of Tensors(MNLFT)model and Nesterov-incorporated Non-negative Latent Factorization of Tensor(NNLFT)model.Empirical studies on two industrial datasets demonstrate that these proposed models can accelerate convergence rate of a learning algorithm,thereby improving model's learning efficiency.