TAROT: Task-Adaptive Refinement of LLM-prior Graphs for Few-shot Tabular Learning

The paper introduces TAROT, a GNN-based framework designed for few-shot tabular learning that enhances predictive performance by constructing and refining a task-adaptive semantic graph. TAROT utilizes a Unified Semantic Tabular Node Encoder (USTNE) to encode heterogeneous tabular data and prompts LLMs to infer semantic relationships, followed by a Task-adaptive Semantic Graph Refinement to optimize the graph structure for specific tasks. This approach addresses the limitations of traditional and LLM-based methods, particularly in managing structural noise and improving feature interaction modeling, thus providing a more efficient solution for practitioners in scenarios with limited labeled data.