Graph-Based Solutions for Explainable and Efficient AI

Abstract

Recent advances in Artificial Intelligence, from Large Language Models (LLMs) to Vision Transformers (ViTs), have demonstrated remarkable capabilities across a variety of tasks. However, these powerful architectures still face critical challenges in terms of explainability, efficiency, and structural understanding. Graph-based modeling offers a unifying framework to address these issues by explicitly representing relationships among data, features, and model components.

This Special Session aims to collect novel contributions on graph-based representations, reasoning, and learning strategies that enhance both the interpretability and computational efficiency of modern AI systems. Graphs provide a natural way to describe structural dependencies, whether among visual tokens in a Transformer, entities in a Knowledge Graph, or Agents in a social network, thus enabling more transparent, robust, and context-aware intelligence.

Submission Information

We invite submissions exploring the integration of graph structures within deep architectures, including but not limited to: graph-based explainability methods for neural networks and Transformers, graph-based methods for fine-tuning of LLMs, graph-driven model pruning and compression, multiplex and heterogeneous graph representations for multimodal data, and graph-based frameworks for social analysis, reasoning, and decision-making.

Past Editions

• Graph-based Solutions for AI @  IJCNN 2025