QMult performance on WN18RR

[adelmemariani]

On WN18RR dataset, I get poor results with QMult. I wonder if there are some specific configurations/hyperparameters that I should use. The QMult works fine with UMLS, KINSHIP, NELL-995-h100 and FB15k-237.

Here are some results for batch size of 1024, learning rate of 0.1, and embedding dimension of 32, with different seeds:

Seed: 1859447115:
"Test": { "H@1": 0.05232929164007658, "H@3": 0.07626037013401404, "H@10": 0.10689215060625398, "MRR": 0.07197849006936502 }

Seed: 694443915:
"Test": { "H@1": 0.05073388640714742, "H@3": 0.07434588385449904, "H@10": 0.11199744735162731, "MRR": 0.07161262733453724 }

Seed: 831769172
"Test": { "H@1": 0.04339502233567326, "H@3": 0.07179323548181238, "H@10": 0.10689215060625398, "MRR": 0.06614106275929006 }

Seed: 2164385946
"Test": { "H@1": 0.047223994894703254, "H@3": 0.07657945118059988, "H@10": 0.11135928525845565, "MRR": 0.06953405514942977 }

Seed: 2376066489
"Test": { "H@1": 0.042437779195915765, "H@3": 0.06955966815571155, "H@10": 0.10338225909380983, "MRR": 0.06371407372612738 }

Seed: 2430986565
"Test": { "H@1": 0.05137204850031908, "H@3": 0.08423739629865985, "H@10": 0.12316528398213146, "MRR": 0.07610544853299074 }