Quality of Experience (QoE) plays a key role in determining the revenue of Internet applications, such as web services and video streaming. While researchers work to improve low-level system performance metrics by using system resources (e.g., network bandwidth and compute cycles) more efficiently, better low-level performance does not always correspond to a better QoE. This is because of users' different QoE sensitivities across different contexts (e.g., application content, user preferences, environments, etc.). Not all low-level system performance improvements are perceptible by users. In this thesis, we introduce a novel user-centric approach to QoE optimization. This approach prioritizes resource allocation based on the dynamic user QoE sensitivities towards performance metrics. However, implementing this approach in Internet applications is not straightforward. Firstly, learning QoE sensitivity can be expensive as it varies significantly across numerous contexts, necessitating this learning for each context. Secondly, such fine-grained resource allocation can introduce substantial system overhead. In this work, we address these challenges and apply our proposed approach to web services and on-demand video streaming. We also introduce a user-study tool to help researchers learn QoE sensitivity more efficiently in terms of monetary cost and latency. Our experiments with real users demonstrate that through this approach, we can either improve user QoE without additional system resource consumption or achieve consistent QoE using fewer resources.