Year: 2024
Type: preprint
Abstract: Graph-based learning models learn structure-aware and node-level representations through relational associations between data points, enhancing predictions and explainability. The ability of Graph Neural Network (GNN) models to learn from non-Euclidean spaces, has shifted explainability efforts towards GNNs, neglecting other significant methodologies. This survey addresses this gap by including traditional machine learning and deep learning models, based on Reinforcement Learning, Multi-Hop Reasoning, Knowledge Graphs, and GNNs. It proposes a hierarchical categorization of graph explanation models based on explanation modalities and approaches. Furthermore, it examines the merits, drawbacks, and suitability of each strategy across domains and tasks followed by future research directions.
Source: HAL
Institutions: Hasso Plattner Institute, Aalborg University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4400968451
