JCSE, vol. 19, no. 1, pp.16-23, 2025

DOI: http://dx.doi.org/10.5626/JCSE.2025.19.1.16

A Federated Learning Method for Multi-Positive and Unlabeled Data Using an Ensemble of One-Positive and Unlabeled Learning Models

Cheong Hee Park
Department of Computer Science and Engineering, Chungnam National University, Daejeon, Korea

Abstract: Centralized learning, which requires data to be collected in a central location, is ill-suited to meet the requirements of stricter regulations on data privacy and security. Federated learning provides solutions to this problem, enabling a model to be trained across a group of clients while maintaining the privacy of each client has small-sized labeled data. When a client has small-sized labeled data from a part of the classes with large-sized unlabeled data, referred to as a multi-positive and unlabeled learning problem, leveraging unlabeled data can improve the performance of the global model aggregated from the local models of the clients. In this paper, we propose a federated learning method for multi-positive and unlabeled data using an ensemble of one-positive and unlabeled learning models. We construct an ensemble of positive and unlabeled learning models for each positive class in clients, which is trained as a binary classifier on one positive class with small-sized labeled data and one negative class consisting of all remaining data. The ensemble collected from all clients is used to extend the labeled data of each class in clients. The experimental results using image and text data show that the proposed method improved the performance for small-sized multi-positive and unlabeled federated learning.

Keyword: No keyword

Full Paper:   8 Downloads, 17 View