Symmetry costs less than the algebra suggests

Published FB15k-237 results, by MRR (mean reciprocal rank of the true entity; higher is better, 1 is perfect):

model	core	MRR
DistMult	superdiagonal	0.343
ComplEx	superdiagonal, complex	0.348
RESCAL	full matrix per relation	0.356
CP-N3	superdiagonal, untied	0.36
ComplEx-N3	superdiagonal, complex	0.37
TuckER	dense, shared	0.358

• TuckER’s own paper listed DistMult at 0.241. The training recipe moved one model by about 0.10 MRR — more than the spread across all six rows. With only about 3% of pairs exposed, symmetry rarely gets the chance to hurt.
• TuckER matches CP-N3 with about 11× fewer parameters. CP-N3’s count is almost all entity tables (8,000 numbers per entity); TuckER’s is mostly its 200³ core, which does not grow with the graph.

Rule out DistMult when entities play both roles in an asymmetric relation; choose a Tucker core when entity count or memory is the constraint.