Asymptotic Evaluation of Classification in the Presence of Label Noise

摘要

In a practical classification problem, there are cases whereincorrect labels are included in training data due to label noise. We introducea classification method in the presence of label noise that idealizes a classificationmethod based on the expectation-maximization (EM) algorithm, andevaluate its performance theoretically. Its performance is asymptoticallyevaluated by assessing the risk function defined as the Kullback-Leiblerdivergence between predictive distribution and true distribution. The resultof this performance evaluation enables a theoretical evaluation of themost successful performance that the EM-based classification method mayachieve.