Evolving Knowledge Graphs for Bias Aware Synthetic Persona Representation

Synthetic persona datasets, such as FinePersonas, are essential for studying AI applications, including large language models (LLMs) and recommender systems. However, conventional storage formats (relational tables, JSON) lack adaptability, limiting real-time updates, deep structural queries, and bias detection. This paper introduces a Dynamic Knowledge Graph (DKG) approach that integrates Graph Neural Networks (GNNs) and Large Language Models (LLMs) to enhance persona representation. Our method enables incremental updates, bias-aware modeling, and efficient querying. Using a large-scale synthetic persona dataset, we demonstrate that DKGs significantly outperform static storage solutions by improving query efficiency (3-8x speedup), adaptive modeling of emergent attributes, and systematic bias mitigation.