arxivst stuff from arxiv that you should probably bookmark

Noise-Tolerant Interactive Learning from Pairwise Comparisons with Near-Minimal Label Complexity

Abstract · Apr 19, 2017 17:00 ·

comparison mellon oracles label pac pairwise labeling oracle email interactive stat-ml

Arxiv Abstract

  • Yichong Xu
  • Hongyang Zhang
  • Kyle Miller
  • Aarti Singh
  • Artur Dubrawski

We study the problem of interactive learning of a classifier from \emph{noisy} labeling and pairwise comparison oracles with near-optimal label complexity. The comparison oracle answers which instance in the given two instances is more likely to be positive, and learning from both oracles has multiple applications ranging from the handwriting recognition to the recommendation system. However, much remains unknown concerning to what extent the comparison oracle helps to improve the label complexity. The challenge is that unlike the conventional PAC-learning framework, the problem of interactive learning with two oracles leverages the two oracles effectively in an adaptive way. In this work, we show that access to the comparison oracle reduces the challenging learning problem to learning of a threshold. We present two plausible algorithms that works well under the Tsybakov and agnostic noise models w.r.t. the comparison and labeling oracles, and show that the overall learning rate can be superior compared with the traditional active learning algorithms in certain cases. Our main results analyze the label and comparison complexity of our algorithms and the costs, benefits, and fundamental limitations of using pairwise comparisons. These results can support multiple applications where preferable form of user feedback is more abstract than a label on each individual instance.

Read the paper (pdf) »