Résumé
Quality of Service (QoS) prediction is an important task in Web service selection and recommendation. Existing approaches to QoS prediction are based on either Content Filtering or Collaborative Filtering. In the two cases, these approaches use external data or past interactions between users and services to predict missing or future QoS scores. One of the most effective techniques for QoS prediction is Matrix Factor-ization (MF), with Latent Factor Models. The key idea of MF consists in learning a compact model for both users and services. Thereafter QoS prediction is simply computed as a dot product between the user's latent model and the service's latent model. Despite the successful results of MF in the recommendation area, there are still a set of problems that should be handled, like: i) the sparsity of the input models, and ii) the learning of the latent factors which is prone to over-fitting. In this paper, we propose an approach to solve these two problems by using a simple neural network, an auto-encoder, and by exploiting cross-validation on a well-known dataset, in order to select the ideal number of latent factors, and thereby reduce the over-fitting phenomenon.