Disentangling decoder capacity and directionality in type constrained knowledge graph link prediction
Abstract
Decoder comparisons in typed knowledge graph (KG) link prediction often confound relation specificity, off-diagonal interactions, and directionality. We separate these factors with four seed-matched decoders applied to GCN, GraphSAGE, and GAT encoders: a shared inner product (IP), a relation-specific diagonal bilinear decoder (RD), a symmetric full bilinear decoder (SB), and an unconstrained directed bilinear decoder (D). The study uses a regulation-grounded coal-mine safety KG with 5,093 nodes and 19,314 canonical directed edges. All 120 decoder–encoder–seed cells use the same ten non-adjacent seeds, type-constrained splits, fixed negatives, and training budget. Two-sided paired AUC tests use pre-specified Holm families, with full filtered MRR and Hits@1/3/10 reported descriptively. RD produced no detectable AUC improvement over IP for any of the three encoders under the matched ten-seed design. SB improved on RD for GraphSAGE (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bf{\Delta}$$\end{document}AUC \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bf{=0.0254}$$\end{document}, Holm \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bf{p=2.15\times10^{-5}}$$\end{document}) and GAT (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bf{\Delta=0.1667}$$\end{document}, Holm \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bf{p=3.24\times10^{-10}}$$\end{document}), but not GCN. Releasing the SB symmetry constraint and enlarging the direction-sensitive score space (D\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$-$$\end{document}SB) increased AUC for GCN (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bf{\Delta=0.0423}$$\end{document}, Holm \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bf{p=4.77\times10^{-5}}$$\end{document}) and GraphSAGE (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bf{\Delta=0.0116}$$\end{document}, Holm \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bf{p=0.00251}$$\end{document}), whereas the GAT contrast was not detected. Thus, no RD-over-IP improvement was detected, while the contributions of symmetric off-diagonal interaction and of releasing symmetry while enlarging the direction-sensitive score space depend on the encoder. A deterministic semantic hard-negative stress test, all-family filtered ranking, and a frozen-embedding symmetric/skew diagnostic support but do not enter the confirmatory tests. Candidate discovery uses threshold-free top-five within-family ranking. These results support a narrow conclusion: decoder function class matters in this typed transductive setting, but no decoder is universally superior.