Abstract:

This paper presents a novel metaheuristic binary crow search algorithm (CSA) designed for positive-unlabeled (PU) learning, a paradigm where only positive and unlabeled data are available, with applications in many diversified fields, such as medical diagnosis and fraud detection. The algorithm represent a useful adaptation of CSA, itself inspired by the food-hiding behavior of crows. The proposed BiCSA-PUL (binary crow search algorithm for positive-unlabeled learning) selects reliable negative samples from unlabeled data using binary vectors, and updates positions employing Hamming distance, guided by a modified F1-score, as fitness function. The algorithm was tested on 30 samples from 10 diverse datasets, outperforming seven state-of-the-art PU learning methods. The results reveal that BiCSA-PUL provides a robust and efficient approach for PU learning, significantly improving fitness and reliability. This work opens new avenues for applying metaheuristic optimization methods to challenging classification problems with limited labeled data. The main limitations are the potentially time-intensive process of parameters tuning and reliance on initial sampling.

Publisher's Version