As the network environment becomes increasingly complex, the threats it faces are becoming more severe. Intrusion detection, as a key proactive defense mechanism in network security, requires more robust and effective detection methods to address these challenges. Graph neural networks (GNNs) have shown excellent performance in anomaly detection. This paper proposes a novel intrusion detection method based on VAE-GNN, an improved graph neural network approach. Specifically, the network traffic data is first constructed into a network traffic graph. This graph is then transformed into a format suitable for processing by the graph neural network to detect anomalous traffic. To address the limitations of traditional intrusion detection methods, such as dataset imbalance and weak feature representation problems, a detection method based on VAE-GNN and the integration of feature statistical importance is proposed. Experimental results on two benchmark intrusion detection datasets demonstrate that the VAE-GNN method outperforms the original graph neural network methods, showing better detection performance and stronger noise resistance.