Abstract
Knowledge graphs (KGs) are vital for enabling knowledge reasoning across various domains. Recent KG reasoning methods that integrate both global and local information have achieved promising results. However, existing methods often suffer from score over-smoothing, which blurs the distinction between correct and incorrect answers and hinders reasoning effectiveness. To address this, we propose DuetGraph, a coarse-to-fine KG reasoning mechanism with dual-pathway global-local fusion. DuetGraph tackles over-smoothing by segregating—rather than stacking—the processing of local (via message passing) and global (via attention) information into two distinct pathways, preventing mutual interference and preserving representational discrimination. In addition, DuetGraph introduces a coarse-to-fine optimization, which partitions entities into high- and low-score subsets. This strategy narrows the candidate space and sharpens the score gap between the two subsets, which alleviates over-smoothing and enhances inference quality. Extensive experiments on various datasets demonstrate that DuetGraph achieves state-of-the-art (SOTA) performance, with up to an 8.7% improvement in reasoning quality and a 1.8× acceleration in training efficiency.
Introduction
Knowledge Graph (KG) reasoning is the fundamental task of predicting missing links within a network of entities and relations. To perform this effectively, models must synthesize two types of information: local neighborhoods (the immediate connections surrounding an entity) and global structures (long-range dependencies and patterns across the entire graph).
While current state-of-the-art models attempt to integrate these signals by stacking message-passing layers (for local context) and transformer layers (for global context), they frequently encounter the problem of score over-smoothing. As these layers are stacked deeper, the mathematical representations of correct and incorrect answers converge, making them nearly indistinguishable. This loss of discriminative ability results in a "blurred" candidate list where the ground-truth answer barely outscores incorrect ones. DuetGraph is designed to overcome this limitation through architectural segregation and a multi-stage optimization strategy.
Methodology
1. Dual-Pathway Global-Local Fusion
DuetGraph addresses over-smoothing by fundamentally changing how local and global information interact. Instead of a sequential stack, DuetGraph employs a parallel architecture:
- The Local Pathway: Utilizes query-aware message passing to capture intricate neighborhood patterns and structural motifs.
- The Global Pathway: Employs an attention mechanism to identify and capture long-range dependencies that local message passing might miss.
These two pathways are processed independently to prevent mutual interference. They are subsequently fused using a learnable parameter (\α), which allows the model to dynamically balance the two signals without "blurring" the underlying entity representations during the propagation phase.
2. Coarse-to-Fine Reasoning Optimization
To further sharpen the distinction between candidates, DuetGraph implements a two-stage inference strategy:
- Stage 1 (Coarse): A lightweight "coarse" model (such as HousE or RED-GNN) performs an initial ranking of all entities. These entities are then partitioned into a high-score "candidate" subset and a low-score "background" subset.
- Stage 2 (Fine): The sophisticated Dual-Pathway model focuses exclusively on refining the scores within these subsets. By narrowing the candidate space, the model avoids the noise of thousands of irrelevant entities.
Mathematically, the authors demonstrate (via Theorem 2) that this coarse-to-fine strategy creates a lower bound on the expected score gap, effectively acting as a "wedge" that forces correct and incorrect answers apart.
Experimental Evaluation
Datasets and Benchmarks
The authors evaluated DuetGraph across a comprehensive suite of benchmarks:
- Inductive Reasoning: FB15k-237, WN18RR, and NELL-995 (testing reasoning on unseen entities).
- Transductive Reasoning: Standard versions of the above plus YAGO3-10.
- Triple Classification: UMLS, FB13, and WN11.
- Large-Scale KGs: Wikidata5M and Freebase.
Key Results
DuetGraph demonstrated significant improvements in both accuracy and computational efficiency:
- Reasoning Quality: The model achieved a state-of-the-art 8.7% improvement in Hits@1 on inductive reasoning tasks, indicating a much higher probability that the top-ranked answer is correct.
- Efficiency Gains: The parallel architecture and coarse-to-fine filtering resulted in a 1.8× acceleration in training efficiency. On large-scale graphs like YAGO3-10, DuetGraph showed a 34.2% increase in inference speed compared to hybrid models like KnowFormer.
- Scalability: On the massive Wikidata5M dataset, DuetGraph outperformed rule-based and embedding-based baselines. Notably, it completed inference in just 80 seconds, compared to over 4,000 seconds for some competing methods.
Discussion and Insights
Segregation vs. Stacking
A primary insight from this research is that "more" layers are not always beneficial in graph neural networks. Stacking layers often leads to a vanishing gradient of information where unique entity features are washed away. By keeping the local and global pathways separate, DuetGraph preserves the "sharpness" of the representations.
Architectural Diversity
Ablation studies revealed that DuetGraph performs optimally when the "coarse" model is architecturally distinct from the "fine" model (e.g., using a triplet-based model to filter for a transformer-based model). This diversity allows the system to capture a wider variety of signals from the data.
Conclusion and Future Work
DuetGraph successfully mitigates the over-smoothing problem in KG reasoning through its dual-pathway fusion and coarse-to-fine optimization. While the model sets new benchmarks for performance and efficiency, the authors note that it remains a "black box" in terms of interpretability. Future research will focus on developing explainability modules to trace reasoning paths through the dual pathways, which is particularly critical for applications in sensitive fields like clinical diagnosis and biomedical research.
Summary updated