抄録
The paper describes new machine learning technique called Transductive Confidence Machine (or Conformal Predictors). The technique is based on recently developed computable approximations of Kolmogorov's algorithmic notion of randomness, and allows us to make reliable predictions using valid measures of confidence in both ""batch"" and ""online"" modes of learning. The advantages are as follows: it can control the number of erroneous predictions by selecting a suitable confidence level; unlike many conventional techniques the approach does not make any additional assumption about the data beyond the iid assumption: the examples are independent and identically distributed; it allows to make estimation of confidence of the prediction for individual examples; can be used as a region predictor with a number of possible predicted values; can be used in high-dimensional problems where number of attributes greatly exceeds the number of objects; it gives well-calibrated predictions that can be used in on-line and off-line learning as well as in ""intermediate"" types of learning e.g. ""slow"", ""lazy"".
