| | Input: Training triples T = {(h, r, t)}, entities E, initial entity matrix WE, relations R, embedding subspace dimension s, dropout probability , candidate sampling rate , regularize parameter α | | Output: subspace adaptive matrix Ws, relation embedding matrix WR, combination operators Deh, Drh, Det, Drt | | Algorithm sub-ProjE (T, E, WE, R, s, , , α) | | (1) | Initializing adaptive matrix Ws, relation matrix WR, combination operators (diagonal matrices) Deh, Drh, Det, Drt with uniform distribution . | | (2) | Loop/A training iteration/epoch/ | | (3) | , , , ;//training data | | (4) | for do/construct training data using all train triples/ | | (5) | | | (6) | if e = = h then/tail is missing/ | | (7) | | | (8) | /all positive tails from T and some sampled negative candidates/ | | (9) | else/head is missing/ | | (10) | | | (11) | /all positive heads from T and some sampled negative candidates/ | | (12) | end if | | (13) | end for | | (14) | for eachdo/minibatches/ | | (15) | | | (16) | for each do/training instance/ | | (17) | | | (18) | | | (19) | | | (20) | end for | | (21) | | | (22) | Update all parameters w.r.t | | (23) | end for | | (24) | EndLoop |
|
