The lack of transparency of graph neural networks (GNNs) poses challenges in understanding the results of e.g., friendship prediction, drug discovery, and community detection. Graph Counterfactual Explanation (GCE) techniques aim to enhance interpretability by generating counterfactual examples, improving trustworthiness, and reducing biases in GNN predictions. However, existing literature on GCE lacks standardization in definitions, methodologies, datasets, and evaluation criteria. To address this, we introduced GRETEL, a comprehensive framework for developing and evaluating GCE methods. GRETEL offers fully extensible built-in components, enabling users to define ad-hoc explanation techniques, generate synthetic datasets, implement customized evaluation metrics, and integrate seamlessly with state-of-the-art prediction models. In this demo, we present GRETEL-2, an enhanced version with a focus on usability and extensibility. We illustrate how these features enhance the interpretability and practicality of GNNs across various scenarios.
GRETEL 2.0: Generation and Evaluation of Graph Counterfactual Explanations Evolved / Prado-Romero, M. A.; Prenkaj, B.; Stilo, G.. - 14948:(2024), pp. 363-367. (Intervento presentato al convegno European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases, ECML PKDD 2024 tenutosi a Vilnius, Lituania) [10.1007/978-3-031-70371-3_21].
GRETEL 2.0: Generation and Evaluation of Graph Counterfactual Explanations Evolved
Prenkaj B.Secondo
Software
;Stilo G.
Ultimo
Supervision
2024
Abstract
The lack of transparency of graph neural networks (GNNs) poses challenges in understanding the results of e.g., friendship prediction, drug discovery, and community detection. Graph Counterfactual Explanation (GCE) techniques aim to enhance interpretability by generating counterfactual examples, improving trustworthiness, and reducing biases in GNN predictions. However, existing literature on GCE lacks standardization in definitions, methodologies, datasets, and evaluation criteria. To address this, we introduced GRETEL, a comprehensive framework for developing and evaluating GCE methods. GRETEL offers fully extensible built-in components, enabling users to define ad-hoc explanation techniques, generate synthetic datasets, implement customized evaluation metrics, and integrate seamlessly with state-of-the-art prediction models. In this demo, we present GRETEL-2, an enhanced version with a focus on usability and extensibility. We illustrate how these features enhance the interpretability and practicality of GNNs across various scenarios.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
