Convex Formulation for Learning from Positive and Unlabeled Data

By Marthinus Christoffel du Plessis, Gang Niu, Masashi Sugiyama. International Conference on Machine Learning, 2015

379 citations recorded by the source corpus.

View Paper: https://www.semanticscholar.org/paper/4ad915a10be9d443206e8ae707d64638be604f8a

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

Citations in the corpus

Loss factorization, weakly supervised learning and label noise robustness (2016, methodology): “For example, labels may be noisy [Natarajan et al., 2013], missing as with semi-supervision [Chapelle et al., 2006] and pu [du Plessis et al., 2015], or aggregated as it happens in multiple instance learning [Dietterich et al., 1997] and learning from label proportions ( llp ) [Quadrianto et al.,…”
Positive-Unlabeled Reward Learning (2019, background): “The key insight of PU learning is that for an appropriately chosen loss function, the true negative risk R g(Dn) can be expressed in terms of positive and unlabeled data as [16, 23] (1− η)R g(Dn) = R g(Du)− ηR g(Dp) .”
Debiased Contrastive Learning (2020, methodology): “Common applications of PU learning are retrieval or outlier detection [ 10 – 12 ].”
Learning from positive and unlabeled data: a survey (2018, methodology): “nlabeled Data: A Survey 25 This general weighting method was proposed in the single-training-set scenario as the rst SAR PU learning method [4] but it already existed before under the SCAR assumption [101,22,49]. The ShiftMC method for matrix completion is also a special case of this method under th”
Positive-Unlabeled Learning with Non-Negative Risk Estimator (2017, methodology): “ u (g) ˇ p; (4) which can be minimized by separating X p and X u with ordinary cost-sensitive learning. An issue is Rb pu(g) in (4) must be non-convex in g, since no ‘(t;y) in (3) can be convex in t. [16] showed that Rb pu(g) in (2) is convex in g, if ‘(t;y) is convex in tand meets a linear-odd cond”
On Symmetric Losses for Learning from Corrupted Labels (2019, background): “Other examples of weakly-supervised learning are learning from positive and unlabeled data [du Plessis et al., 2014, 2015; Kiryo et al., 2017], learning from pairwise similarity and unlabeled data [Bao et al., 2018], and learning from complementary labels [Ishida et al., 2017a].”
Weakly supervised learning via statistical sufficiency (2016, background): “For example, labels may be missing as with semi-supervision [Chapelle et al., 2006] and positive and unlabeled data [du Plessis et al., 2015], noisy [Natarajan et al., 2013], aggregated as it happens in multiple instance learning [Dietterich et al., 1997] and learning from label proportions [Kuck…”
Federated Learning with Only Positive Labels (2020, background): “Diﬀerent from PU learning (Liu et al., 2002; Elkan & Noto, 2008; Plessis et al., 2015; Hsieh et al., 2015), in the federated learning setting, the clients do not have access to unlabeled data for both positive and negative classes.”