Adaptive K-NN metric classification based on improved Kepler optimization algorithm

K-nearest neighbor (K-NN) method has been widely utilized in data mining and pattern recognition due to its elegant geometric basis and well-defined statistical characteristics. The nearest neighbor number K and the distance metric employed by K-NN have a great impact on the classification performance. To effectively improve the K-NN classification capacity, this article proposes an improved Kepler optimization algorithm (termed IKOA) with the interaction effect renewal mechanism and the dynamic rebalancing steady state mechanism for enhancing the global exploration and local exploitation. Furthermore, the IKOA-KNN algorithm, a novel K-NN variant, is constructed by integrating IKOA and K-NN for reinforcing the classification accuracy of the canonical K-NN. In the classification process, the proposed IKOA is used to optimize the adaptive K value of K-NN and the well-posed p -norm value of the distance metric, synchronously. In addition, feature selection based on IKOA is used to eliminate irrelevant and redundant features, so as to maintain or enhance classification accuracy. The experimental results of IKOA on CEC-2017 benchmark functions show that the proposed algorithm, compared with seven well-known meta-heuristic algorithms (i.e., WSO, BWO, DMOA, FLA, RSA, SPO, KOA), the proposed algorithm enhances the exploitation and exploration ability, improves the convergence speed and is more stable when solving optimization problems. Fourteen bi-and multi-class datasets from the UCI Machine Learning Repository are used to demonstrate the effectiveness of the proposed IKOA-KNN in addressing real-world optimization problems. The statistical results of IKOA-KNN validate its excellent classification accuracy with up to 100% on some of these datasets.