The pseudonymous nature of blockchain introduces security risks through abnormal accounts, which can undermine system stability and compromise user asset security. To address this, we propose a graph learning method guided by correlation sampling for detecting abnormal accounts. Our approach constructs a transaction graph enriched with both node and relationship features. A novel correlation-based strategy is employed to calculate correlation coefficients and sample relevant neighbors. A graph neural network is then used to learn node representations, with an attention mechanism dynamically adjusting neighbor influence. Experimental results demonstrate that our method outperforms existing approaches in detecting abnormal accounts.