Recipe-Controlled Decoder Audit for Structural Knowledge-Graph Completion

A new audit protocol challenges how researchers attribute performance gains in knowledge-graph completion, arguing that decoder choice—often overlooked—can be the main diagnostic, especially on small datasets [1]. The recipe-controlled decoder audit (RCDA) evaluates structural transductive knowledge-graph completion by swapping decoders while holding the training recipe constant [1]. The primary controlled pair consists of the ComplEx and DistMult models, with targeted spot-checks using RotatE and TransE [2]. Across five standard knowledge graphs, the ComplEx-versus-DistMult differences under the controlled recipe are modest but consistent, ranging from +0.005 to +0.012 mean reciprocal rank (MRR) [2]. Encoder effects, such as those from CompGCN-style architectures, vary more by dataset [2]. On small knowledge graphs, decoder effects become the primary diagnostic [1]. The Kinship dataset shows a stable ComplEx advantage of +0.143 MRR across six seeds [2]. On the UMLS dataset, ComplEx holds a +0.022 MRR advantage in a clean six-seed server rerun, but the result reverses in an earlier provenance variant [2]. The authors therefore treat small-graph decoder choice as recipe- and provenance-sensitive rather than as a fixed dataset winner [2]. The audit further reveals that decoder choice interacts with encoder depth on the WN18RR dataset [1]. Under the controlled recipe, a depth-zero ComplEx configuration on YAGO3-10 reaches 0.6971 ± 0.0048 MRR at dimension 128 [2]. The protocol recommends three reporting practices: publish matched decoder rows, log small-graph provenance, and sweep decoder-by-depth combinations before making encoder-level claims [2]. Knowledge-graph completion research has historically focused on encoder innovations, with decoder selection often treated as a secondary implementation detail. The RCDA framework formalizes a counterweight to that tendency by isolating the decoder’s contribution through controlled experimentation [1]. The work appears as a preprint on arXiv and has not yet been peer-reviewed [1].