Deep forest auto-Encoder for resource-Centric attributes graph embedding

Abstract

Graph embedding is an important technique used for representing graph structure data that preserves intrinsic features in a low-dimensional space suitable for graph-based applications. Graphs containing node attributes and weighted links are commonly employed to model various real-world problems and issues in computer science. In recent years, a hot research topic has been the exploitation of diverse information, including node attributes and topological semantic information, in graph embedding. However, due to limitations in deep learning based on neural networks, such information has not been fully utilized nor adequately integrated in existing models, leaving graph embedding unsatisfactory, especially for large resource graphs (e.g., knowledge graphs and task interaction graphs). In this study, we introduce a resource-centric graph embedding approach based on deep random forests learning, which reconstructs graphs using a deep autoencoder to achieve high effectiveness. To accomplish this, our approach employs three key components. The first component is a preprocessor driven by graph similarity, alongside modularity and self-attention modules, to comprehensively integrate graph representation. The second component utilizes local graph information structures to enhance the raw graph. Finally, we integrate diverse information using multi-grained scanning and dual-level cascade forests in the deep learning extractor and generator, ultimately producing the final graph embedding. Experimental results on seven real-world scenarios show that our approach outperforms state-of-the-art embedding methods.