Instance-Dependent Positive and Unlabeled Learning With Labeling Bias Estimation

Chen Gong, Qizhou Wang, Tongliang Liu, Bo Han, Jane You, Jian Yang, Dacheng Tao

Abstract

This paper studies instance-dependent <u>P</u> ositive and <u>U</u> nlabeled (PU) classification, where whether a positive example will be labeled (indicated by <inline-formula><tex-math notation="LaTeX">$s$</tex-math></inline-formula> ) is not only related to the class label <inline-formula><tex-math notation="LaTeX">$y$</tex-math></inline-formula> , but also depends on the observation <inline-formula><tex-math notation="LaTeX">$\mathbf {x}$</tex-math></inline-formula> . Therefore, the labeling probability on positive examples is not uniform as previous works assumed, but is biased to some simple or critical data points. To depict the above dependency relationship, a graphical model is built in this paper which further leads to a maximization problem on the induced likelihood function regarding <inline-formula><tex-math notation="LaTeX">$P(s,y|\mathbf {x})$</tex-math></inline-formula> . By utilizing the well-known EM and Adam optimization techniques, the labeling probability of any positive example <inline-formula><tex-math notation="LaTeX">$P(s=1|y=1,\mathbf {x})$</tex-math></inline-formula> as well as the classifier induced by <inline-formula><tex-math notation="LaTeX">$P(y|\mathbf {x})$</tex-math></inline-formula> can be acquired. Theoretically, we prove that the critical solution always exists, and is locally unique for linear model if some sufficient conditions are met. Moreover, we upper bound the generalization error for both linear logistic and non-linear network instantiations of our algorithm, with the convergence rate of expected risk to empirical risk as <inline-formula><tex-math notation="LaTeX">$\mathcal {O}(1/\sqrt{k}+1/\sqrt{n-k}+1/\sqrt{n})$</tex-math></inline-formula> ( <inline-formula><tex-math notation="LaTeX">$k$</tex-math></inline-formula> and <inline-formula><tex-math notation="LaTeX">$n$</tex-math></inline-formula> are the sizes of positive set and the entire training set, respectively). Empirically, we compare our method with state-of-the-art instance-independent and instance-dependent PU algorithms on a wide range of synthetic, benchmark and real-world datasets, and the experimental results firmly demonstrate the advantage of the proposed method over the existing PU approaches.
BibTeX
@article{gong2022instance,
  author = {Chen Gong and Qizhou Wang and Tongliang Liu and Bo Han and Jane You and Jian Yang and Dacheng Tao},
  title = {{Instance-Dependent Positive and Unlabeled Learning With Labeling Bias Estimation}},
  journal = {IEEE Trans. Pattern Anal. Mach. Intell.},
  year = {2022},
  url = {https://openreview.net/forum?id=NgfH-36RBFv},
}

Loading related content...

Figure