Enhancing graph structure learning through multiple features and graphs fusion

In recent years, Graph Structure Learning (GSL) methods have attracted considerable attention, due to their ability to optimize and clean the graph structure from noise. Most of the available GSL methods focus on learning a single graph. However, incorporating multiple graph structures can significantly enhance the robustness and generalization capacity of a model by effectively capturing diverse and multifaceted knowledge. Therefore, there has recently been a notable shift in attention towards the multi-GSL. These methods either deal with the fusion of the learned graphs or the fusion of the learned feature obtained from these graphs. However, very limited studies have explored the fusion of both learned graphs and features. To cope with these issues, this paper introduces MFGSL (multiple feature graph structure learning), a novel approach that simultaneously learns and merges multiple graph structures and features to learn an informative and comprehensive graph structure. The effectiveness of our proposed method is assessed through extensive experiments on five benchmark datasets: Cora, Citeseer, PubMed, Amazone photo, and Digits. Our results demonstrate the superiority of our proposed method over other state-of-the-art methods.