<p>Unsupervised anomaly detection is crucial for monitoring and ensuring the reliability and security of cloud systems. However, this task is challenging due to the complexity and high dimensionality of cloud system data. We observe that cloud system metrics demonstrate strong inter-correlations. These correlations establish a reliable foundation for anomaly detection, with deviations providing clear signals of potential anomalies. Existing methods fail to fully exploit these correlations for anomaly detection. To address this challenge, we propose CTGAD(Correlation-based Transformer and Graph Anomaly Detector), a novel model that integrates GCN and Transformer to leverage both structural and sequential information. Based on the observation, we compute the data correlations and embed them into the adjacency matrix of the GCN, enriching the network with valuable prior information to enhance its anomaly detection capabilities. Furthermore, we leverage the concept of correlation to improve the attention mechanism, proposing a novel Corre-attention that dynamically selects the most relevant keys for each query based on correlation scores. This mechanism effectively minimizes unnecessary distractions in attention allocation of Transformer. Extensive experiments on five public datasets show that CTGAD achieves the best F1-scores on JSS1 and MBD and the highest macro-average F1-score across all five datasets, reaching 49.25%.</p>