Deep Manifold Attributed Graph Embedding (DMAGE) is a novel graph embedding framework that aims to tackle the challenge of unsupervised attributed graph representation learning, which requires both structural and feature information to be represented in the latent space. Existing methods can face issues with oversmoothing and cannot directly optimize representation, thus limiting their applications in downstream tasks. In this article, we will discuss the DMAGE framework and how it can be used to surpass state-of-the-art methods on three popular datasets.
Background
Graph representation learning is an essential task that involves mapping nodes in a graph to a lower-dimensional latent space to facilitate downstream machine learning tasks such as visualization, clustering, and link prediction. However, existing methods have limitations when dealing with attributed graphs, which contain both structural and feature information. Additionally, these methods typically rely on reconstruction tasks that can lead to oversmoothing, which reduces the discriminative power of the learned representations. Therefore, developing a method that can effectively learn low-dimensional representations of attributed graphs without oversmoothing is crucial.
The DMAGE Framework
The DMAGE framework proposes a novel approach to graph embedding that addresses the limitations of existing methods. It uses a node-to-node geodesic similarity to compute the inter-node similarity between the data space and the latent space, and then uses Bergman divergence as a loss function to minimize the difference between them. Additionally, DMAGE incorporates graph structure augmentation to improve representation stability and employs a network structure with fewer aggregation to alleviate the oversmoothing problem.
Results
DMAGE has been extensively evaluated on three popular datasets, namely Cora, Citeseer, and Pubmed. The framework has been compared to state-of-the-art methods, and it has surpassed them by a significant margin on unsupervised visualization, node clustering, and link prediction tasks. These results demonstrate the effectiveness of DMAGE as a graph embedding framework.
The DMAGE framework provides a novel approach to unsupervised attributed graph representation learning that addresses some of the limitations of existing methods. With its use of geodesic similarity and Bergman divergence as loss functions, DMAGE can optimize representations more directly and prevent oversmoothing. Additionally, the incorporation of graph structure augmentation and the use of a network structure with fewer aggregations further improve the representation stability. DMAGE exceeds state-of-the-art methods on popular datasets, demonstrating its potential as an effective approach to graph embedding.
