Abstract
Quality of Experience (QoE) consists of a set of indicators that show the perceived satisfaction of using a multimedia (or other kind of) service by the end user. Being so the QoE presents a subjective metric and the only relevant mechanisms for measuring such indicators are subjective tests. Due to the fact that subjective tests are an expensive, impractical and in cases of live streaming a close to impossible exercise we set out on a twofold task to address this issue. First we set out to build prediction models using traditional Machine Learning (ML) techniques based on subjective test data. Second we explore an approach for reduction of the training dataset that will minimize the need for subjective data whilst keeping the prediction models as accurate as possible. For the first goal we used supervised learning based classification algorithms and we came up with high accuracy (over ninety percent) for the prediction models. To address the issue of high cost training data we developed a novel approach in reducing the training dataset while keeping a high accuracy of the classifiers. The reduction method provides a grading mechanism for unseen data. By having this mechanism in the online learning platform we can optimize the process of asking for user feedback by looking for the most significant cases, and therefore improving the gain on the trade-off between more feedback and more accuracy.