• Trade-off between Accuracy and Labeling Cost

    • Aim for high accuracy and low labeling cost.
  • Various means of weakly supervised learning (mostly research involving Professor Masaru Sugiyama)

    • When there are data for the positive class “blue” and “unknown (blue/red)”, divide them into blue and red.
    • Classification based on confidence: Divide into blue and red using only the data with the probability of being blue.
      • Effective in environments with only successful data or biased data, etc.
    • Classification based on similarity data pairs: Classify using only the information “similar to data X” and unlabeled data.
    • Classification based on complementary labels (labels indicating that X is not class A) only. (Expert in information science) #MachineLearning