Weighted Logistic Regression Modeling for Semi-Supervised Classification

• Activity Number: 295
• Type: Contributed
• Date/Time: Tuesday, July 31, 2012 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Shuichi Kawano*+
• Companies: Osaka Prefecture University
• Keywords: Covariate shift ; Model selection ; Regularization ; Semi-supervised learning

Abstract

Semi-supervised learning, which is a statistical method that combines both labeled and unlabeled data, has received considerable attention in contemporary statistics and machine learning. There are a lot of modeling approaches for semi-supervised learning; e.g., a mixture modeling approach, a graph theory approach, a boosting approach, a logistic regression modeling approach and so on. Most of these semi-supervised approaches implicitly assume that a density function for labeled data is the same as that for unlabeled data. In this talk, we propose a weighted logistic model for semi-supervised classification problem by using covariate shift adaptation techniques in the situation that the density function for labeled data is different from that for unlabeled data. A crucial issue in modeling process is the choice of adjusted parameters included in the models. In order to select the adjusted parameters, we introduce an information criterion that evaluates our semi-supervised logistic models. Numerical examples are given to examine the effectiveness of proposed semi-supervised discrimination.