A Boosting Algorithm for Training from Only Unlabeled Data

• Published in: Advanced Data Mining and Applications Vol. 13726; pp. 459 - 473
• Main Authors: Zhao, Yawen, Yue, Lin, Xu, Miao
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer 2022; Springer Nature Switzerland
• Series: Lecture Notes in Computer Science
• Subjects: Boosting; UU learning; Weakly supervised learning
• ISBN: 3031221362; 9783031221361
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-031-22137-8_34

Unlabeled-unlabeled (UU) learning was proposed to cope with the high cost of data annotation and some realistic cases, in which we cannot get labeled data. It allows us to train a classifier with only unlabeled data. State-of-the-art (SOTA) UU methods with good performance based on neural networks (NN) have been proposed; however, there is a lack of studies on boosting algorithms for learning from only unlabeled data, even though boosting algorithms sometimes perform very well with simple base classifiers. We propose a novel boosting algorithm for UU learning: Ada-UU, which compares against neural networks. The proposed method follows the general procedure of AdaBoost while the classification error is estimated with two sets of unlabeled (U) data. We empirically demonstrate that Ada-UU outperforms neural networks on several large-scale benchmark UU datasets and has comparable performance on a small-scale benchmark dataset.