Abstract
Multi-label classification becomes especially difficult for sparse, noisy labels and complex relationships among them. False flag labeling modifies the labels of training instances to construct a new training set that can improve model performance. An evolutionary algorithm treats this relabeling as an optimisation problem without assuming that added or removed labels correct annotation errors. Experiments across 50 datasets and seven classification models demonstrate the approach’s effectiveness.
Citation
García-Pedrajas, N. E., Cuevas-Muñoz, J. M., Mendoza-Hurtado, M., Pérez-Rodríguez, J., & de Haro-García, A. (2026). False flag: An evolutionary false labeling approach for multilabel classification. Applied Soft Computing, 194, Article 114912. https://doi.org/10.1016/j.asoc.2026.114912