Community detection is a crucial field in graph data mining, gaining attention due to graph neural networks (GNNs). Graph autoencoder network (GAE) has shown potential in unsupervised community detection. However, GAE-based methods are often complex with multiple layers and parameters. They also overlook neighbor node distribution disparities in imbalanced graph data. In an effort to overcome these constraints, we propose a self-supervised graph autoencoder network with node feature convolution (SGAE-NFC) for community detection. SGAE-NFC reduces model complexity through a node feature convolution layer. A self-supervised model optimizes representation learning and community detection, enhancing robustness. Experimental results on real-world datasets demonstrate SGAE-NFC’s superior performance, surpassing existing methods in accuracy and robustness.