Relational GCN (R-GCN)

Real knowledge bases are not simple graphs: edges carry relation types (born-in, works-for), and a plain GCN cannot tell them apart. R-GCN extends graph convolution to multi-relational data by giving each relation type its own transformation matrix, with basis-decomposition and block-diagonal regularization to keep parameter counts manageable when there are hundreds of relations. The model is used two ways: as a standalone encoder for entity classification, and as an encoder feeding a factorization decoder (DistMult) for link prediction, accumulating evidence over multiple inference steps in the graph.\n\nThe ESWC 2018 paper reports that the R-GCN + DistMult combination beats the DistMult-only baseline on FB15k (filtered MRR 0.696 vs 0.634) and WN18, and that the encoder gives a 29.8% improvement on FB15k-237 over a decoder-only baseline. The authors are candid that LinkFeat, a simple baseline exploiting inverse relation pairs, still wins on the older FB15k/WN18 splits. R-GCN became the canonical neural encoder for knowledge graphs and the starting point for CompGCN and later multi-relational architectures.

Paper-reported: R-GCN+DistMult filtered MRR 0.696 on FB15k vs 0.634 DistMult-only; 29.8% gain on FB15k-237 over decoder-only baseline (Schlichtkrull et al., ESWC 2018). LinkFeat still beats it on FB15k/WN18, which the authors acknowledge.