Personality detection is widely used in personalized recommendation and mental health assessment. The widespread use of social media provides abundant user-generated text data for automatic personality analysis. Recently, Graph Neural Network (GNN)-based methods have attracted attention, but two major challenges remain. The first problem is that social media texts often reflect temporary emotions rather than stable personality traits. The second problem is that severe class imbalance in datasets weakens predictive performance for minority classes. To address these issues, we propose the Residual-weighted Dynamic Graph Convolutional Network (R-DGCN), introducing an adaptive residual weighting mechanism that dynamically adjusts information transfer between nodes to better capture stable personality features. We also design a weighted loss function to mitigate class imbalance. The proposed R-DGCN surpasses current models on Pandora and Kaggle, indicating its robustness in mitigating noise and class imbalance issues in graph-based personality prediction.