Label Distribution Learning with Discriminative Instance Mapping

Heng-Ru Zhang, Run-Ting Bai, Wen-Tao Tang

Abstract: Label distribution learning (LDL) is an effective tool to tackle label ambiguity since it allows one instance to be associated with multiple labels in different degrees. Therefore, the more complex but informative label space makes it challenging to directly model the relationship between original features and label distributions. In this paper, an algorithm called Label Distribution Learning with Discriminative Instance Mapping (LDLDIM) is proposed to select a discriminative instance pool (DIP) to map the original features into a more discriminative space. First, we design a criterion that incorporates label information to quantify the discriminative power of each instance. Second, we select several instances with the highest discriminative ability to construct the DIP, and map the instances to the discriminative space through the DIP. By exploiting label information, this criterion enables the selected DIP to ensure that instances that are close (far away) in label space remain close (far away) in the discriminative space. Finally, multiple regressions for prediction are trained on the label distributions and the new features that are obtained by distance mapping with DIP. Experiments and comparisons on 16 datasets illustrate that our algorithm outperforms 6 state-of-the-art LDL methods in most cases.