SynergyKGC: Reconciling Topological Heterogeneity in Knowledge Graph Completion via Topology-Aware Synergy

TL;DR

SynergyKGC addresses the structural resolution mismatch in Knowledge Graph Completion by fusing semantic signals with topology via an instruction-driven, dual-tower architecture. A Semantic Expert establishes a robust semantic manifold, while a Synergy Expert, guided by density-aware Identity Anchoring and Cross-Modal Synergy Attention, adaptively retrieves and fuses topological context, with a Dynamic Dual-Tower Consistency mechanism ensuring training and inference remain aligned. The method uses a two-phase training regime and joint optimization with L_NCE and L_align to achieve rapid convergence and high predictive precision, demonstrated by state-of-the-art results on FB15k-237 and especially a remarkable +8.0% absolute gain in Hits@1 on WN18RR. This framework offers a scalable principle for resilient information integration in non-homogeneous graphs, balancing structural sufficiency and identity redundancy across densities, and significantly reducing preprocessing and warm-up overhead.

Abstract

Knowledge Graph Completion (KGC) fundamentally hinges on the coherent fusion of pre-trained entity semantics with heterogeneous topological structures to facilitate robust relational reasoning. However, existing paradigms encounter a critical "structural resolution mismatch," failing to reconcile divergent representational demands across varying graph densities, which precipitates structural noise interference in dense clusters and catastrophic representation collapse in sparse regions. We present SynergyKGC, an adaptive framework that advances traditional neighbor aggregation to an active Cross-Modal Synergy Expert via relation-aware cross-attention and semantic-intent-driven gating. By coupling a density-dependent Identity Anchoring strategy with a Double-tower Coherent Consistency architecture, SynergyKGC effectively reconciles topological heterogeneity while ensuring representational stability across training and inference phases. Systematic evaluations on two public benchmarks validate the superiority of our method in significantly boosting KGC hit rates, providing empirical evidence for a generalized principle of resilient information integration in non-homogeneous structured data.

Figures (5)

• Figure 1: The overall architecture of SynergyKGC. The framework establishes a progressive cross-modal reconciliation between latent semantic manifolds and explicit topological signals, where initial textual embeddings ($e^{sem}$) from the Semantic Expert are enhanced by the Synergy Expert via an asymmetric dual-tower design: the $(h,r)$ stream employs Relation-Aware Cross-Attention for context retrieval, while the $t$ stream utilizes a Relation-Agnostic counterpart to maintain representation consistency during inference. This integration is governed by a density-adaptive identity anchoring strategy that dynamically toggles entity-self signals based on a dataset-specific threshold $\phi$ ($\phi=1$ for dense FB15k-237; no threshold for sparse WN18RR) to prevent representation collapse while suppressing structural noise, all within a joint optimization framework where adaptive gating ($\alpha$) is constrained by contrastive ($\mathcal{L}_{NCE}$) and MSE alignment ($\mathcal{L}_{align}$) objectives to effectively mitigate semantic drift during structural-semantic fusion.
• Figure 3: Analysis of synergy activation thresholds (WN18RR). The heatmaps illustrate the performance landscape across varied start epochs (vertical axis) and total training epochs (horizontal axis). Red boxes highlight the global optima across the entire grid, while bold values denote the peak performance within each column, indicating the optimal activation timing for a given training budget. Notably, activation at Epoch 20 consistently yields the most stable and superior results across all metrics, justifying its selection as the default threshold.
• Figure 4: Ablation analysis of the proposed Dual-Tower Synergy Mechanism. Left: Evaluation of Architectural Synergy reveals that the structural consistency between the Query and Entity towers is the prerequisite for aligning head-tail semantic manifolds. Right: Investigation into Lifecycle Synergy demonstrates that keeping the synergy mechanism consistent across training and inference stages is critical for mitigating the distribution shift, thereby ensuring that learned collaborative features are fully activated during deployment.
• Figure 6: Structural sufficiency in FB15k-237. For 50 Cent + award_nominee, dual bridge entities (Dr. Dre, Eminem) form a closed relational path between high-degree nodes, enabling precise localization independent of identity anchoring and validating the "Structure $\approx$ Identity" phenomenon.
• Figure 7: Popularity bias and structural scaffolding in WN18RR. For the query life_science + hypernym, the sparse target (degree 3) and the hub-like distractor (maths, degree 73) are topologically indistinguishable via a shared generic bridge. The Identity Anchoring strategy serves as a critical scaffold to overcome this "degree trap."