Query Performance Prediction (QPP) is a critical task in Information Retrieval (IR), enabling the estimation of query effectiveness without requiring relevance judgments. Traditional QPP methods, however, struggle to consistently capture the contextual nuances within and between queries, similar queries, and retrieved documents, often relying on limited, surface-level representations. This limitation hinders their ability to handle diverse query types and retrieval scenarios, resulting in variable performance across datasets. To address these challenges, this PhD research proposal introduces a novel framework for QPP based on contextualized representations, enhancing both adaptability and robustness. This proposal explores three primary research directions to build a comprehensive, context-aware approach to QPP. The first research direction focuses on term-level analysis by assigning difficulty weights to individual terms using fine-tuned language models. This term-specific approach identifies elements that either enhance or impede query effectiveness, providing a precise understanding of query composition. Building on this foundation, the second research direction introduces graph-based contextual representations using disturbance generation to evaluate query robustness through adaptive semantic perturbations in the embedding space. This method leverages metrics from node, edge, and cluster properties to enable robust predictions across datasets. Finally, this PhD research proposal aims to advance QPP by incorporating supervised knowledge graphs with Graph Neural Networks (GNNs), enhancing scalability and accuracy across diverse retrieval scenarios. This work will contribute to the field by delivering a framework that captures complex query relationships, improving prediction reliability across varied IR settings.