A novel Multi-Label Leaming Approach for Missing and Noisy Labels
Xiaolei Yang, Kai Liu, Shenming Gu, Anhui Tan · 2023
Recently, many scholars have shown a strong interest in weak-supervised multi-label learning, to solving which idealizing label information is a common approach. However, in real life, we cannot obtain the complete labels of instances. In addition, the obtained labels may contain noise, which can reduce the performance of the algorithms. To address the above issues, we introduce MLML, a novel multi-label learning method which is designed specifically to handle missing labels and noisy labels simultaneously. Our method involves the following several steps. Firstly, we utilize the global information of multi-label data to construct a label correlation matrix. Through this matrix, we can better understand the association and influence between different labels. Secondly, in order to mitigate the impact of noisy labels, we introduce label popularity normalization, which improves the performance of the algorithm by considering the frequency of labels in the data. Finally, we improve the method by integrating positive and negative label information. We also consider the relationship between feature similarity and label similarity, which ensures that our method not only considers the information of a single label, but also the mutual influence between labels and the association between features and labels. Based on extensive experiments conducted on multiple datasets, our algorithm MLML has demonstrated comparable performance to state-of-the-art algorithms.