The influence of social media platforms has fundamentally transformed modern communication paradigms while introducing new security challenges. While these platforms offer users rich social connections and interactive experiences, malicious social bots are also covertly exploiting their unique features to engage in harmful activities, such as spreading misinformation, manipulating public opinion, and disrupting the digital ecosystem. However, conventional detection methods relying on static feature engineering or homogeneous network analysis prove inadequate against continuously evolving bot mimicry techniques and imbalanced dataset attacks. To address these limitations, this work proposes a hierarchical detection framework integrating heterogeneous graph representation learning with dynamic sample balancing strategies. By constructing multi-dimensional interaction graphs to model user behavior and fusing them with similarity-based affinity graphs, our approach enables graph neural networks to identify latent bot behavioral patterns effectively. Furthermore, the hybrid sampling mechanism rebalances sample distributions to mitigate data imbalance. Evaluations on real-world datasets verify the effectiveness of our proposed method. Our study not only advances methodological innovation in computational social science but also provides theoretical foundations for platform governance and digital policy-making.