With the amount of global network traffic steadily increasing, mainly due to video streaming services, network operators are faced with the challenge of efficiently managing their resources while meeting customer demands and expectations. A prerequisite for such Quality-of-Experience–driven (QoE) network traffic management is the monitoring and inference of application-level performance in terms of video Key Performance Indicators (KPIs) that directly influence end-user QoE. Given the persistent adoption of end-to-end encryption, operators lack direct insights into video quality metrics such as start-up delays, resolutions, or stalling events, which are needed to adequately estimate QoE and drive resource management decisions. This research has been motivated by the challenge to devise an approach for the estimation of QoE/KPIs from encrypted traffic, where we recognised machine learning (ML) methods as a promising way forward, and a fundamental part of the methodology. In this thesis, we present a generic methodology for training of ML–based models for the estimation of QoE and application-level KPIs of adaptive video streaming services, applicable in the context of traffic encryption. The methodology is embodied in the form of a conceptual framework which demonstrates the key methodological steps involved in developing an in-network QoE/KPI monitoring solution, including model training, model deployment, and model re-evaluation and adaptation. The methodology is evaluated through six studies, conducted over the course of four years, involving YouTube video on demand, resulting in models for session-level QoE/KPI estimation focused on both Android and iOS, models for near real-time KPI estimation, analysis of cross-platform and cross-network model applicability, analysis of methods for automated model re-evaluation and adaptation, and the analysis of the impact of the inclusion of application-level data (possibly shared by a service provider) on the performance of QoE/KPI estimation models. The key contribution of the thesis is a methodology that identifies relevant KPIs to be used as prediction targets, identifies relevant network traffic features obtainable on IP-level, and describes the procedures of model training, evaluation, re-evaluation, and adaptation, thus covering processes that are a prerequisite for actual model deployment. The practical focus of the thesis has been on YouTube, which is one of the most prominent video streaming services today. In that context, a unique and valuable contribution are also the models for YouTube QoE/KPI estimation focused on mobile platforms, applicable for both TCP and QUIC traffic, and employing a standardised ITU-T P.1203 model for the calculation of QoE. Moreover, the thesis presents models that include application-level context data as additional predictors, thus contributing to motivation for resolving existing issues standing in the way to QoE-centric cooperation among actors involved in the service delivery chain.