End-to-End Graph Flattening Method for Large Language Models

TL;DR

Inspired by human cognitive reasoning habits, a novel method for graph flattening to fit LLMs is proposed, termed as End-to-End DAG-Path prompting (EEDP), which enhances the reasoning performance of LLMs in long-distance scenarios while maintaining excellent performance in short-distance scenarios.

Abstract

In recent years, the breakthrough of Large Language Models (LLMs) offers new ideas for achieving universal methods on graph data. The common practice of converting graphs into natural language for LLMs, which refers to graph flattening, exhibits good generalizability and interpretability. However, the poor organization of the textual format results in poor performance in long-distance scenario understanding. Inspired by human cognitive reasoning habits, we propose a novel method for graph flattening to fit LLMs, termed as End-to-End DAG-Path prompting (EEDP). Experiments on real-world datasets show that EEDP enhances the reasoning performance of LLMs in long-distance scenarios while maintaining excellent performance in short-distance scenarios, demonstrating good robustness in the face of distance variations.

Figures (1)

• Figure 1: The framework of EEDP and an example of how EEDP handles a input graph $\mathcal{G}$. It is first used to generate $DAG_{\text{EEDP}}$. $Path_{\text{EEDP}}$ are extracted from $\mathcal{G}$ based on the endpoints $\mathcal{V}_{\text{end}}$ found by using $DAG_{\text{EEDP}}$. $Path_{\text{EEDP}}$ can be compressed if needed. Finally, $Path_{\text{EEDP}}$ and $\mathcal{G}_{\text{adjlst}}$ are concatenated to obtain the EEDP-flattened graph $\mathcal{G}_{\text{EEDP}}$.