Is architectural complexity overrated? Competitive and interpretable knowledge graph completion with relate

Abhijit Chakraborty, Chahana Dahal, Ashutosh Balasubramaniam, Tejas Anvekar, Vivek Gupta

Keywords: Knowledge Graph Completion (Link Prediction), Knowledge Graph Embeddings (KGE), Real‑valued Phase–Modulus Decomposition, Generative Probabilistic Circuits / Energy‑based Modeling, Robustness to Structural Perturbations, Interpretability & Attribution, Computational Efficiency (Training/Inference), Benchmarks: FB15k‑237, WN18RR, YAGO3‑10, ogbl‑biokg

TL;DR: RELATE is a simple, real‑valued KGE with a phase–modulus design and a generative probabilistic view that achieves competitive/SOTA link prediction (incl. ogbl‑biokg, YAGO3‑10), strong robustness, and lower compute than complex baselines.

Abstract:

We introduce RELATE, a minimal, fully real–valued and interpretable KGE whose phase–modulus decomposition cleanly separates directional semantics (a smooth sinusoidal phase for (anti-)symmetry and composition) from magnitude/type semantics (a slope–weighted modulus with a lightweight type bias). We prove full expressivity for finite KGs and establish invariance/identifiability and training–stability
results. Despite its simplicity, RELATE admits a tractable energy and a smooth, decomposable probabilistic–circuit compilation that enables likelihood training and exact conditional sampling, bridging discriminative link prediction with generative
reasoning. On ogbl-biokg (typed protocol) and FB15k–237 / WN18RR / YAGO3–10, RELATE matches or exceeds strong baselines at matched capacity (including SOTA on YAGO3–10) while using ≈24% less training compute, achieving ≈31% higher
throughput, and ≈22% lower peak memory. Robustness studies under counterfactual injection, relation swaps, and edge edits show markedly smaller MRR drops than TransE/RotatE, and ablations isolate the roles of phase vs. modulus and type
bias. Overall, RELATE shows that carefully structured geometry, rather than added depth, drives accuracy, efficiency, and interpretability.We also analyze the Z-paradox failure mode of distance-based KGE and show reduced confounding on Z-Hard splits.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 22313