Monitoring of Quality of Service (QoS) in high-speed Internet infrastructure is a challenging task. However, precise assessments must take into account the fact that the requirements for the given quality level are service-dependent. Backbone QoS monitoring and analysis requires processing of large amount of the data and knowledge of which kind of application the traffic belongs to. To overcome the drawbacks of existing methods for traffic classification we proposed and evaluated a centralized solution based on C5.0 Machine Learning Algorithm (MLA) and decision rules. The first task was to collect and provide C5.0 high-quality training data, divided into groups corresponding to different types of applications. It was found that currently existing means of collecting data (classification by ports, Deep Packet Inspection, statistical classification, public data sources) are not sufficient and they do not comply with the required standards. To collect training data a new system was developed, in which the major role is performed by volunteers. Client applications installed on their computers collect the detailed data about each flow passing through the network interface, together with the application name taken from the description of system sockets. This paper proposes a new method for measuring the Quality of Service (QoS) level in broadband networks, based on our Volunteer-Based System for collecting the training data, Machine Learning Algorithms for generating the classification rules and application-specific rules for assessing the QoS level. We combine both passive and active monitoring technologies. The paper evaluates different implementation possibilities, presents the current implementation of particular parts of the system, their initial runs and obtained results, highlighting parts relevant from the QoS point of view.