Label Distribution Changing Learning with Sample Space Expanding
Authors: Chao Xu, Hong Tao, Jing Zhang, Dewen Hu, Chenping Hou; 24(36):1−48, 2023.
Abstract
As data collection methods continue to evolve, label ambiguity has become prevalent in various applications. The challenge lies in reducing this uncertainty and leveraging its effectiveness. Two types of representative label ambiguities, Label Distribution Learning (LDL) and Emerging New Class (ENC), have received significant attention. However, in certain applications like emotion distribution recognition and facial age estimation, a more complex label ambiguity scenario arises, namely label distribution changing with sample space expanding due to the introduction of new classes. This paper addresses this rarely studied problem by proposing a new framework called Label Distribution Changing Learning (LDCL), along with its theoretical guarantee of generalization error bound. Our approach expands the sample space by re-scaling previous distributions and estimates the emerging label value using a scaling constraint factor. We present two special cases within the LDCL framework, along with their optimizations and convergence analyses. In addition to evaluating LDCL on 13 existing datasets, we also apply it to the application of emotion distribution recognition. Experimental results demonstrate the effectiveness of our approach in tackling label ambiguity and estimating facial emotion.
[abs]