www-ai.cs.tu-dortmund.de/LEHRE/SEMINARE/WS2122/TrustworthyAIMachineLearning/weiss2018a.pdf
10 50 100 1000 Train 50 30, 30, 30 11,11,155 99.9 99.8 99.9 99.9 99.9 100 11.0 11,10,11 100 99.9 99.9 99.9 100 500 30, 30, 30 10,10,10 100 99.9 100 99.9 100.0
1:1 ratio between the positive and negative [...] and so AR,p should satisfy ˆδR,p(w1) 6= ˆδR,p(w2). The R-states h1 = ĝ(w1) and h2 = ĝ(w2) are passed, along with p, to ref , to yield a new, finer, partitioning p′ for which ˆδR,p′(w1) 6= ˆδR,p′(w2).
This [...] satisfying:
1. ∀h1 ∈ H , p′(h) 6= p′(h1), and
2. ∀h1, h2 ∈ S, p(h1) 6= p(h2)⇒ p′(h1) 6= p′(h2).
(In practice, condition 1 may be relaxed to separating at least one of the vectors in H from h, and our method …