Personalized Quality Of Experience Evaluation In Mobile Internet Based Onmachinelearning

The development of mobile communication technology, Internet technology and smart end device urges the spring up of mobile internet services, which is an indispensable part of peoples’ everyday life nowadays. And the mobile Internet will keep the tendency of high-speed development in the future. An important support for this development is the good Quality of Experience (QoE) for the users. The research on QoE evaluation attracts a lot of researchers, who have accomplished many excellent achievement. But most of these studies are only takes the objective factors into consideration, that are easy to observe or to measure. Few studies have addressed the relationship between QoE and users’ subjective factors, thus limiting the evaluation of QoE to the statistic of user groups. And it is hard to refine the granularity of evaluation to per-user per-service. Specific to this research status, this paper proposes a personalized QoE (PQoE) evaluation and studies the corresponding approaches. The research contents include the following.Firstly, a data collecting platform is built to collect the data relating to the personalized QoE. This platform can offer the online video service to the users, while gathering the necessary data in the background. Those data are the foundation of the PQoE study.Secondly, this thesis studies the relationship between QoE and user preference, which is one of the most important subjective features. Three properties are concluded based on the related researches and reasonable assumptions, and then the mathematic model are built conforming to the properties. This model takes the user subjective features into consideration, distinguishing itself from traditional models, and can embody the personalized difference between users. The core part of the PQoE evaluation model is users’ specific preference towards services. In practical applications, this preference cannot be achieved by users’feedback towards all the services, so the prediction of this preference is necessary, that need to use machine learning technologies. After studying different prediction model and algorithms, a Bayesian Graphic Model (BGM) is proposed, containing three layers, that is, observed features layer, latent features layer and prediction layer from bottom to up. Due to the factor that the training of this model has no analytical solution, the Expectation Maximization algorithm is adopted. And the Monte Carlo sampling is used to simplify the step of calculating expectation. Then the well-known MovieLens dataset is used in this thesis to verify this model, as well as analyze the parameters in this model. The experiment results show that the performance of BGM is much better than that of Matrix Factorization (MF), with13%decrease in terms of the Root Mean Square Error (RMSE).Considering the Restricted Boltzmann Machine (RBM) is capable to extract the latent features well, a collaborative Filtering model is built based on RBM. The standard training algorithm, Contrastive Divergence (CD), is adopted. To relieve the cold-start problem as well as improve the performance, this thesis extends the model by subjoining users’observed features as extra visible nodes. Experiments are conducted to analyze the parameters of the model and to compare with the extended model. The result shows that the extend model has a notable lower RMSE than the original one.