Node Embedding Enhancement Model Based on Joint Optimization of Clustering Distribution

Deep graph clustering methods combining Autoencoder (AE) and Graph Neural Networks (GNNs) have accomplished excellent performance. These methods utilize attributes and graph structure in the encoding phase, but only reconstruct the graph structure in the decoding phase, thereby limiting the utilization of attributes. Besides, these methods do not consider the consistency of the clustering distributions learned by AE and GNN, which provides incorrect supervisory information for the iterative optimization of the model. Accordingly, this paper proposes a node embedding enhancement model based on joint optimization of clustering distribution (NEE-JOD), which clusters nodes based on the joint optimization of AE and Graph Auto-Encoder (GAE). First, the node embedding enhancement mechanism is designed by combining AE and GAE to enhance the node embedding, and then the dual decoders are trained to reconstruct the attributes and graph structure simultaneously. Besides, the joint optimization mechanism for clustering distributions is introduced, which imposes consistency constraints by comparing the clustering distributions generated by the AE module and the GAE, thus generating more reliable clustering distributions and providing correct supervisory information for the iterative optimization of the model. Experiments on five datasets validate the effectiveness of the proposed model.