Detection of mislabeled training data in pattern recognition with influence function(Session 2b)

概要

論文の詳細を見る
Sensitivity analysis based on influence functions has been widely studied in the field of statistics. In particular the evaluation approach has been applied to different statistical methods such as principal component analysis, correspondence analysis, and linear discriminant analysis. However, the study of discriminant methods in pattern recognition is less advanced. With this background, we focused on a subspace method, which is a discriminant method in pattern recognition, and proposed an evaluation method for the influence of training samples to the result of analysis using influence functions. However, the performance and effectiveness of our method were not illustrated well. In this study, we focused on our single-case diagnostics and applied the approach to a representative subspace method, following which we showed good results. Specifically, in situations that had mislabeled samples in the training data, we were able to detect such samples using our approach and subsequently deleted them from the training data to enhance the performance of the target classifier.
2011-11-11