Bat algorithm for variable selection in multivariate classification modeling using linear discriminant analysis

Variable selection is an efficient and powerful tool for reducing the dimensionality of multivariate data and multicollinearity, enabling the successful classification of samples by Linear Discriminant Analysis (LDA). This paper describes a bat-inspired algorithm as an alternative to performing variable selection in multivariate classification by LDA. Named BA-LDA, this algorithm simulates the echolocation behavior of bats when moving in search of prey. It was implemented with a cost function associated with the average risk of misclassification in LDA. The performance of BA-LDA was evaluated on mass spectrometry (MS), near-infrared (NIR), and ultraviolet-visible (UV-vis) spectrometric data sets of serum from unaffected and affected women with ovarian cancer, coffee, and vegetable oil samples, respectively. Its performance was compared with the genetic algorithm (GA-LDA) and successive projection algorithm (SPA-LDA). As the main results, BA-LDA presented a classification performance similar to the GALDA and SPA-LDA, classifying all coffee and vegetable oil samples. For the ovarian cancer dataset, BA-LDA (93% accuracy) presented a better classification performance than GA-LDA (88% accuracy) and SPA-LDA (79% accuracy). Regarding the stochastic nature and reproducibility, the BA-LDA algorithm tends to select variables in regions associated with chemical information and with better separation between classes.

Automated summary

This paper introduces a bat-inspired algorithm for variable selection in linear discriminant analysis (LDA). The algorithm simulates the echolocation behavior of bats when searching for prey and uses a cost function associated with the average risk of misclassification in LDA. The results show that the algorithm performs similarly to genetic algorithm (GA-LDA) and successive projection algorithm (SPA-LDA) in classifying coffee and vegetable oil samples, and outperforms them in classifying ovarian cancer samples.