An intelligent feature selection approach based on moth flame optimization for medical diagnosis

Abstract

In this work, an enhanced moth flame optimization (MFO) algorithm is proposed as a search strategy within a wrapper feature selection (FS) framework. It aims mainly to improve the classification tasks in medical applications. FS is an NP-hard problem because the run time of its procedure grows exponentially. The persistent necessity for a powerful FS system and the promising results of swarming behavior in various optimization scenarios motivated our attempts to develop a new FS approach. This paper uses a recently developed methodology inspired by the moth movement style in searching for a near-optimal feature subset for reliable disease diagnosis. The proposed modification strategy is based on two stages of enhancement. In the first stage, eight binary variants are produced using eight transfer functions. In the second stage, the Levy flight operator is integrated into the structure of MFO in combination with transfer functions. The main target is to increase the diversity of the algorithm and support the exploration of the search space. Twenty-three medical data sets downloaded from UCI, Keel, Kaggle data repositories are used to validate the proposed approaches. It has been demonstrated that the proposed approach significantly outperforms other well-known wrapper approaches across 83% of data sets. Furthermore, the proposed approach outperforms other methods in the literature across 75% of the data sets. The comparisons with filter-based approach reveal superior performance across 70% of the data sets. This work also conducts an extensive investigation of the parameters and studies the effects of different settings on the performance of the FS process. The empirical results and various comparisons reveal a remarkable effect of the Levy flight operator and transfer functions on the performance of MFO. This can be achieved through improving the diversity, achieving a greater exploration/exploitation trade-off, and enhancing the convergence trends of the optimizer.