Estimation de la QoS dans les services Web par apprentissage profond.

Abstract

In this thesis, we propose a deep learning-based approach wich combines a matrix factorization model based on a deep auto-encoder (DAE) and a clustering technique. Three variants of the auto-encoder design have been used. The first one is composed of a single hidden layer that represents the vector of latent factors of users and/or services. A second architecture considers several hidden layers. A third model consists of a combination of a deep auto-encoder model and a generative adversarial network. Other problems underlying the estimation of missing QoS values were addressed in this work. The first one is related to the vulnerability of prediction systems to the data sparsity problem. To deal with this issue our proposal consists of in using a clustering algorithm based on Kohonen’s self-organising maps, where the initialization is done using location attributes. The second one that we have dealt with is the cold start problem, which occurs when adding new users/services to the prediction system. The latter one is globally managed by exploiting a spatial features as well. The conducted experiments show that our proposals can provide better performances in terms of QoS prediction, and consequently provide more guidance for users in their choice of preferred services than existing methods do. The QoS parameters on which we relied on to carry out our various experiments are response time and throughput. However, the proposed QoS prediction algorithms can be applied to other QoS factors.