Decide link to record as machine learning compliance restriction
I have pairs of sets like
A = { L, M, N, P } = { <"Lll", 47, 0.004>, <"Mm", 60, 0.95>, <"Nnnn", 33, 0.2892>, <"P", 47, 0.0125> }
B = { l, m, n, o } = { <"l", 46, 0.004>, <"m", 0, 0.95>, <"nn", 33, 0.2892>, <"oOo", 33, 0.5773> }
... and I want to automatically train the algorithm based on well-known data in order to know how to relate the elements of a set as
{ <L, l>, <M, m>, <N, n>, <?, o>, <P, ?> }
... with at most one match for each element of any set. Sets should not have the same size and have no guarantees regarding their coincidence - perhaps there are no matches, maybe all matches, maybe a combination of matches and inconsistencies. But in many cases, compatibility with people is expected, and the computer should come close to it.
Tried so far
H(a, b, w1, w2, w3)evaluates a pair of tuples <a1, a2, a3>from Aand <b1, b2, b3>from Bhow f1(a1, b1) * w1 + f2(a2, b2) * w2 + f3(a3, b3) * w3, where f1, f2and are f3processed manually and w1, w2and w3are parameterized weights. I sort all pairs A Γ Bwith my ratings and take pairs for which none of the participants is already represented by a pair with a higher score. I use rough hill climbing to train the scales so that the resulting pairs are displayed in accordance with the training data. The ideal weighting configuration has a threshold tthat determines the correct pair estimates S_abfrom the incorrect pair estimates. This algorithm usually finds ideal configurations after several hundred or thousands of iterations for my training data of about 800(A, B)sets containing 2500 pairs of 8-uples (instead of the 3-upl illustrated). I have not yet given him a set of validation data to find out how much this method processes.
I am not happy with the hard-worked aspect of the problem. I can only imagine machine learning methods for pairs of glasses, but the subsequent mapping is done manually and maybe not as smart as the ideal solution, taking into account the set-display as a whole. Since part of machine learning does not take into account the whole set, it seems to me that I am missing some information that he can use to make better decisions.
, , A Γ B S_ab = < f1(a1, b1), f2(a2, b2), ..., fn(an, bn) > ( n-), [n, ?, 1] S_ab. /, , .
, , , , ||A|| ||B|| . , A B , . , ?
...
, ? . , . , . - , , .