Abstract
Selecting a single global neighbourhood size is a major limitation of multi-label k-nearest-neighbour methods because different regions of the feature space contain different label distributions and decision boundaries. This work associates a potentially different k value with every prototype and optimises it from the local effect of neighbouring candidates. Experiments across 20 problems show significant improvements over standard and locally adaptive multi-label k-nearest-neighbour baselines while retaining comparable test-time complexity.
Citation
Romero-del-Castillo, J. A., Mendoza-Hurtado, M., Ortiz-Boyer, D., & García-Pedrajas, N. (2022). Local-based k values for multi-label k-nearest neighbors rule. Engineering Applications of Artificial Intelligence, 116, Article 105487. https://doi.org/10.1016/j.engappai.2022.105487