Generic Embedded Semantic Dictionary for Robust Multi-Label Classification
Zhengming Ding, Ming Shao, Sheng Li, Yun Fu · 2018
Multi-label classification has attracted great attention in various applications and generated significant interest in data mining and learning fields. For the incompleteness of multi-label data, numerous approaches were developed to address partially missing labels in multi-label data, and traditional multi-label algorithms mainly adopt low-rank embedding and graph regularizer to recover the missing labels. However, how to simultaneously approach missing labels and discriminant multi-label embedding within the low-rank regime is still unclear. In this work, we propose a Generic Embedded Semantic Dictionary (GESD) learning framework for robust multi-label classification, where we both consider the partially and totally missing labels for the visual data. Specifically, we explore a low-rank coding strategy to encode visual features with recovered label matrix by constructing an effective semantic dictionary. In this way, the low-rankness will be appropriately propagated to recover multi-labels and improve label correlation, given missing labels in the training stage. Extensive experiments on six real-world benchmarks verify that our method can correctly capture label correlation and achieve better label recovery & prediction results than the state-of-the-art algorithms.