A hybrid framework integrating graph neural networks and large language models for automated data visualization

Abstract

Data visualization is crucial for exploratory data analysis; however, automating this process remains challenging due to the complexity of semantic reasoning and syntactic code generation. To address this, this study introduces a hybrid framework that integrates Graph Neural Networks (GNN) and Large Language Models (LLM). The proposed pipeline first analyzes the attribute relationships within the dataset using a trained GNN model to identify the most informative feature pairs. Subsequently, it utilizes a Google Flan-T5 model - fine-tuned on a comprehensive dataset of 21,207 examples - to generate valid Vega-Lite visualization specifications. The training process achieved robust convergence with a training loss of 0.0309 and a validation loss of 0.00131. Experimental results across 10 diverse public datasets confirm that the system achieves a 100% syntactic validity rate, with Top-1 and Top-3 accuracies reaching 66% and 100%, respectively. These results demonstrate that the framework produces meaningful and interpretable visual representations while significantly reducing manual workload and human bias in visualization design.