Distributed Semisupervised Partial Label Learning Over Networks
Ying Liu, Zhen Xu, Chen Zhang · IEEE Transactions on Artificial Intelligence · 2022
Partial label learning (PLL) deals with the classification from sufficient training data associated with a candidate set of labels but not the only correct one. In this article, we focus on PLL with some ambiguously labeled and many unlabeled data collected from multiple nodes distributed over a network. To solve this problem, a distributed semisupervised PLL (dS$^2$PLL) framework is formulated, in which the weighted logistic loss with respect to partially labeled (PL) and unlabeled data is utilized. In the proposed dS$^2$PLL algorithm, the parameters of the classifier are adapted in a collaborative manner, meanwhile the weights of each training sample and the ground-truth confidence of candidate labels are adaptively learned to disambiguate the correct label from the candidate label set. The performance of the proposed algorithm is analyzed theoretically and verified by simulations on both synthetic and real datasets. Results show that the proposed algorithm achieves good classification performance and robustness to the ambiguity in the labels.