Graph Neural Network (GNN) is a powerful method for intrusion detection in IoT networks, offering higher detection performance compared to traditional models. However, its black-box nature challenges trustworthiness, as the decision-making process is not transparent. Additionally, the lack of explainability hinders error diagnosis and targeted performance improvement. To address this, we propose a two-stage explanation-driven GNN model that enhances both interpretability and detection performance for IoT network intrusion detection. In the first stage, class-wise feature importance is extracted using post-hoc Explainable AI (XAI) methods applied to an E-GraphSAGE model. In the second stage, the insights are used to refine the feature space, and the E-GraphSAGE model is reused with an updated classification layer following a One-vs-All (OvA) strategy to enhance multi-class detection. The experiments used publicly available IoT network traffic datasets and results demonstrated that the proposed OvA-based E-GraphSAGE model enhanced the overall detection rate and the F1-score. The class-wise explainability not only enhanced the model transparency but also enabled targeted improvements in prediction accuracy. The implementation of our work is available at https://github.com/EANimesha/TrustGuard