RCGELM-AE: an efficient graph embedding deep model without gradient descent

Abstract

Graph embedding encodes high-dimensional sparse graph data into low-dimensional features for downstream tasks. Existing Graph Autoencoders (GAEs) can integrate structural and attribute information by graph convolution, but their training relies on gradient descent often results in slow convergence and susceptibility to local optima. To address above challenges, this study proposes a model called the Residual compensation graph Convolutional Generalized Extreme Learning Machine Autoencoder (RCGELM-AE), which is built on the Extreme Learning Machine Autoencoder (ELM-AE) and incorporates two key designs. Firstly, it embeds graph convolution operations between the input and hidden layers, compensating for GELM-AE’s inherent inability to capture the topological information of graphs. Secondly, it introduces a residual compensation mechanism that extracts deep-level features from the reconstruction errors of the encoder. This mechanism not only avoids the over-smoothing risk caused by repeatedly stacked graph propagation commonly observed in deep graph models, but also enhances the model’s representational power. Specifically, in link prediction tasks, AUC and AP are improved by 4.11%–8.93% and 4.01%–10.01%. In node classification tasks, F1-micro and F1-macro scores increase by 0.42%–4.97% and 0.48%–5.27%, respectively, with a notable reduction in running time. Overall, this framework offers a novel, gradient-descent-free solution for achieving efficient and effective graph embedding.