核心概念
本文提出了一種名為 G-SPARC 的新型譜架構,旨在解決圖學習中冷啟動節點的預測問題,特別是在節點分類、節點分群和鏈路預測等任務中。
統計資料
Cora 數據集包含 2,708 個節點、5,429 條邊緣和 7 個類別。
Citeseer 數據集包含 3,312 個節點、4,732 條邊緣和 6 個類別。
Pubmed 數據集包含 19,717 個節點、44,338 條邊緣和 3 個類別。
Reddit 數據集包含 232,965 個節點、11,606,919 條邊緣和 41 個類別。
引述
"Graphs have advanced deep learning techniques across various domains, enabling tasks such as node classification, node clustering, and link prediction."
"A major, yet often overlooked, challenge in graph learning is generalizing to nodes that emerge without initial connections."
"To address the limitations of traditional graph learning methods, we leverage a fundamental concept in graph theory by transitioning from a graph representation defined by the adjacency matrix to its spectral representation captured through the eigenvectors of the Laplacian matrix."