CADReN: Contextual Anchor-Driven Relational Network for Controllable Cross-Graphs Node Importance Estimation

TL;DR

This work tackles cross-graph Node Importance Estimation (NIE) for Retriever-Augmented Generation by introducing CADReN, a Contextual Anchor-driven architecture that computes node relevance relative to user-defined anchors. It combines semantic and structural encoders, CA–BG cross-attention, reconstruction-based robustness, and an attention-based aggregation, followed by a post-processing step that fuses initial predictions with semantic and structural similarity signals. CADReN demonstrates strong cross-graph performance with zero-shot transfer, while preserving competitive single-graph NIE results, and introduces open datasets RIC200 and WK1K to spur cross-graph NIE research. The approach yields flexible, controllable NIE aligned with user interests, with practical impact for KG-enhanced LLM retrieval and decision support, and suggests directions for richer structural encodings and better CA–node distance modeling. $I_{final} = \sigma(\alpha I_{init} + \beta \mathcal{S}_{sem} + \gamma \mathcal{S}_{str})$ and $\ abla\mathcal{L}_{total} = \nabla\Big(\mathcal{B}(I_{gt}, I_{final}) + \mathcal{L}_{sem} + \mathcal{L}_{str}\Big)$ express core training dynamics, while CA enables adaptable retrieval across diverse graphs.$

Abstract

Node Importance Estimation (NIE) is crucial for integrating external information into Large Language Models through Retriever-Augmented Generation. Traditional methods, focusing on static, single-graph characteristics, lack adaptability to new graphs and user-specific requirements. CADReN, our proposed method, addresses these limitations by introducing a Contextual Anchor (CA) mechanism. This approach enables the network to assess node importance relative to the CA, considering both structural and semantic features within Knowledge Graphs (KGs). Extensive experiments show that CADReN achieves better performance in cross-graph NIE task, with zero-shot prediction ability. CADReN is also proven to match the performance of previous models on single-graph NIE task. Additionally, we introduce and opensource two new datasets, RIC200 and WK1K, specifically designed for cross-graph NIE research, providing a valuable resource for future developments in this domain.

Figures (6)

• Figure 1: CADReN leverages user-defined Contextual Anchors (CAs) to enhance precision in KG queries. In the figure, KG-Apple contains diverse information related to Apple. Users applying Company-Tech and Fruit-Tree CAs receive focused outputs via CADReN, contrasting with the generalized results given by previous NIE networks without CA utilization."
• Figure 2: The figure above presents the overall architecture of the CADReN model. The semantic and structural information in CA and BG are encoded in BERT and our proposed structural encoder, respectively. Cross-attention fusion is then applied to the combinations of these embeddings to capture the relational information between CA and BG. The BG embeddings mixed with the information from CA are then used to predict the NIE scores, with the introduction of Reconstruction Auto-encoder, Attention-based Aggregation mechanism and Post-Processing mechanism to improve the quality of the output.
• Figure 3: Attention based Aggregation mechanism. The Aggregation matrix contains trainable attention parameters, which are used to produce the self-attention Query that guides the prediction of Node Importance Score.
• Figure 4: Top 20 nodes with highest NIS predicted. Red (resp. orange) nodes are GT nodes corresponding to CA 1 (resp. CA 2) nodes.
• Figure 5: Results of experiment on BG No. 1608708