Target Learning: A Novel Framework to Mine Significant Dependencies for Unlabeled Data

Abstract

To mine significant dependencies among predictiveattributes, much work has been carried out to learn Bayesian netwrok classifiers (BNC$_\mathcal {T}$s) from labeled training data set $\mathcal {T}$. However, if BNC$_\mathcal {T}$ does not capture the “right” dependencies that would be most relevant to unlabeled testing instance, that will result in performance degradation. To address this issue we propose a novel framework, called target learning, that takes each unlabeled testing instance as a target and builds an “unstable” Bayesian model BNC$_\mathcal {P}$ for it. To make BNC$_\mathcal {P}$ and BNC$_\mathcal {T}$ complementary to each other and work efficiently in combination, the same learning strategy is applied to build them. Experimental comparison on 32 large data sets from UCI machine learning repository shows that, for BNCs with different degrees of dependence target learning always helps improve the generalization performance with minimal additional computation.