Core Concepts
CoDy outperforms GreeDy in generating clear explanations for TGNN decision-making, increasing transparency and trustworthiness.
Abstract
この論文は、動的グラフにおけるTGNNのための新しい対事実的説明手法であるGreeDyとCoDyを紹介しています。CoDyは、Monte Carlo Tree Searchを活用してGreeDyを上回り、TGNNの意思決定に関する明確な説明を生成し、透明性と信頼性を高めます。論文では、選択戦略が説明パフォーマンスに大きな影響を与えることも示されています。
Directory:
Introduction
GNNs have been successful in various areas like drug discovery and weather forecasting.
Dynamic graphs are more representative of real-world scenarios.
Related Work
Various explanation methods exist for GNNs on static graphs.
Preliminaries and Problem Formulation
Explaining future link predictions involves forecasting upcoming edge addition events based on past events in the temporal graph.
Proposed Methods: GreeDy and CoDy
Search space constraints, search tree structure, selection strategies, and algorithms for both methods are discussed.
Experiments
Evaluation metrics include fidelity, sparsity, runtime, search iterations, and Jaccard similarity across different datasets and models.
Stats
CoDyはGreeDyと既存の事実的手法よりも優れた結果を示しました。
CoDyは最大59%の成功率で有意な対事実的入力を見つけることができます。
Quotes
"Counterfactual explanations offer actionable insights, help identify biases, and can uncover potential adversarial examples."
"CoDy stands out as a highly efficient method for generating counterfactual explanations for TGNN models."