具幾何感知能力之圖對比學習框架

Abstract

由於圖神經網路（Graph Neural Networks, GNNs）在實際應用中常面臨標註資料不足的問題，因此採用自監督式學習（self-supervised learning）來訓練圖神經網路已成為一項重要課題。其中，對比學習（contrastive learning）作為非監督式學習中的核心方法，已被廣泛應用於圖神經網路領域。然而，對比學習的效能高度依賴於正樣本與負樣本的選擇方式，如何合理定義正負樣本，一直是對比學習中的關鍵挑戰。 本研究提出 GeoGCL，一個具備幾何感知能力的圖對比學習框架，將幾何先驗資訊引入正樣本的生成與負樣本的加權機制中。不同於以往方法僅依賴隨機或可學的增強策略，GeoGCL 利用預測出的方向與距離比例，在嵌入空間中對錨點進行幾何變換，以生成保留結構語義的正樣本。為了避免生成樣本與原始圖結構偏離過大，我們引入一個基於圖自編碼器（Graph Autoencoder）的重建模組作為正則項，約束生成樣本的拓撲一致性。此外，GeoGCL 依據負樣本與錨點之間的角度與距離關係進行分類，並根據其困難度與誤標潛在性自適應地重新加權對比損失。實驗結果顯示，GeoGCL 在多個基準數據集上均達到優異的準確率，展示了在圖對比學習中納入幾何資訊的重要性。

Graph Neural Networks (GNNs) often suffer from a lack of labeled data in real-world scenarios, making self-supervised learning an essential approach for training them. Among various unsupervised methods, contrastive learning has gained significant attention and has been widely adopted in the GNN domain. However, the effectiveness of contrastive learning largely depends on how positive and negative samples are defined—a critical challenge in its application. We propose GeoGCL, a novel Geometry-aware Graph Contrastive Learning framework that incorporates geometric priors into both the generation of positive samples and the reweighting of negative ones. Unlike previous methods that rely on randomized or learned augmentations without considering embedding geometry, GeoGCL generates structure-preserving positive pairs by perturbing the anchor embedding along a learned direction and distance, guided by predicted angular and radial scaling. To ensure semantic and structural fidelity, we introduce a Graph Autoencoder-based reconstructor as a regularization component, which encourages the generated positives to remain topologically consistent with the original graph. Furthermore, GeoGCL classifies negative samples based on their angular and distance proximity to the anchor, and adaptively reweights their contribution in the contrastive loss to better model hard and false negatives. Extensive experiments on benchmark datasets demonstrate that GeoGCL achieves superior performance, showcasing the importance of geometric awareness in graph contrastive learning.