Interactive bi-space embedding based temporal graph learning framework

As a fundamental branch of representation learning, temporal graph representation learning has garnered considerable attention from both academic researchers and industry pratitioners. Although existing temporal graph learning methods show good performances, these methods generally embed temporal graph into a single representation space which cannot simultaneously model time-domain features and structure-domain features well. Inevitably, existing works obtain a suboptimal temporal graph representation which further degrades the performance of downstream analysis tasks. To overcome above problem, we theoretically derive a novel I nteractive B i-Space E mbedding based T emporal G raph L earning F ramework named IBE-TGLF. Specifically, based on feature distribution discrepancy, IBE-TGLF employs two embedding spaces to modeling dynamic graph evolution and temporal graph structure. Through contrastive space alignment, IBE-TGLF can alleviate the semantic gap between different spaces and yield high-performance graph representation. Finally, we conduct extensive experiments on multiple benchmarks to validate the effectiveness and efficiency of IBE-TGLF. Experimental results demonstrate the state-of-the-art performances of IBE-TGLF.