基於多重訊號之圖對比學習偽陰性選擇

Abstract

圖對比學習 (Graph Contrastive Learning, GCL) 已成為圖資料自監督表示學習的重要方法。在節點層級的 InfoNCE 式圖對比學習中，模型通常將同一節點在不同增強視圖中的表示視為正樣本，並將其他節點視為負樣本。然而，許多被視為負樣本的節點對實際上可能屬於相同類別，形成偽陰性(false negative) 問題。這些錯誤的負樣本會對同類節點施加排斥力，可能影響表示空間與節點分類表現。 現有偽陰性處理方法大多從當前嵌入相似度或靜態資訊來判斷節點對關係。這些方法提供了重要且有效的線索，但較少有研究系統性地比較不同類型的訊號是否能找出不同的偽陰性候選，以及這些訊號在組合後是否具有互補性。本論文因此將偽陰性選擇分成三類無標籤訊號：embedding-based、structural、以及 temporal training dynamics，並比較三者的選擇品質與互補性。 我們設計多種節點對分數，包括 embedding similarity、neighborhood overlap、drawdown、trajectory variance 等。實驗分析每個訊號的偽陰性選擇品質，並進一步比較 Union、Vote 與 Intersect 三種 multi-signal 組合策略在不同選擇數量下的行為。結果顯示 embedding 與 structural 訊號通常能提供較可靠的候選，而 temporal 訊號能提供少量獨特但較不穩定的資訊。三種組合策略呈現不同的 precision、coverage 與 downstream performance 取捨。在合適的選擇數量下，本論文提出的多訊號選擇框架能改善多個 GCL 模型的節點分類表現，並可作為靈活的 plug-and-play 元件。

Graph Contrastive Learning (GCL) has become an important method for self-supervised representation learning on graph-structured data. In node-level InfoNCE-based GCL, the same node under two augmented views is treated as a positive pair, while other nodes are treated as negatives. However, some node pairs treated as negatives may actually belong to the same class, causing the false negative problem. These false negatives introduce repulsive force between same-class nodes and may affect the learned representation space and downstream node classification performance. Most existing false-negative handling methods rely on current embedding similarity or other static information to estimate the relation between node pairs. These signals are useful, but few studies systematically examine whether different types of label-free evidence can identify different false-negative candidates, and whether these signals are complementary when combined. This thesis therefore studies false-negative selection from three families of label-free signals: embedding-based signals, structural signals, and temporal training-dynamics signals. We design several pair-level scores, including current embedding similarity, neighborhood overlap, drawdown, and trajectory variance. We first analyze the false-negative selection quality of each signal and then compare Union, Vote, and Intersect strategies under different selection budgets. The results show that embedding-based and structural signals usually provide more reliable candidates, while temporal signals provide distinct but less stable information. The three strategies show different trade-offs among precision, coverage, and downstream performance. With suitable budgets, the proposed multi-signal selection framework improves multiple GCL baselines and can be used as a flexible plug-and-play component for graph contrastive learning.