This thesis explored YouTube's performance when there was some user interaction being performed in the iFrame YouTube player of a previously developed YouQ Android application. To do so, most of the previously developed YouQ system was redeveloped and the test methodology has been adapted accordingly. In the Android application, network traffic capture was implemented, scripts for data processing have been changed and developed, application for calculating network features has been changed and new datasets were collected in a laboratory environment for testing the new models. The first part of this thesis focuses on used user interaction scenarios and newly implemented changes on the YouQ system. A number of datasets were created in order to analyse the impact of user interaction on certain level of YouTube performance. The first thing that was analyzed is how user interaction impacts application-level data. What was concluded is that some interactions cause certain events (i.e., paused event, buffering event) in the player and produce stallings which have direct impact on the QoE classification. Analysis of buffer state shows how some user interactions can easily by detected and how in some cases the API for buffer information returns false information. The next thing that was analysed is the network-layer data. Most interesting was how to detect user interaction in that layer so new prediction models could be applied on them. Analysis of captured network traffic was done in Wireshark with all its features such as IO graphs and conversation statistics. Results showed that there is no reliable way to detect user interaction from captured network traffic. The final part of the thesis was creation of one large big dataset which consists of almost all datasets used throughout the thesis and of 445 YouTube videos. Main goal of this part was to create a machine learning classification model using the WEKA tool for five different algorithms. Results have shown how new model has an accuracy of around 90% with RandomForest being the algorithm with the highest accuracy. Dataset was also tested using the previously built YouQ system model (built without user interaction scenarios) to compare results. The comparison has shown that classification accuracy can be improved when building a new model involving user interaction scenarios. However, it should be noted that the training dataset did not have a balanced number of classified videos per class. For that reason, ongoing and future work will focus on collecting a larger and balanced dataset, and further validating developed models.