Application of genetic algorithm (GA) to select input variables in support vector machine (SVM) for analyzing the occurrence of roach, Rutilus rutilus, in streams

Abstract

Support vector machine (SVM) was used to analyze the occurrence of roach in Flemish stream basins (Belgium). Several habitat and physico?chemical variables were used as inputs for the model development. The biotic variable merely consisted of abundance data which was used for predicting presence/absence of roach. Genetic algorithm (GA) was combined with SVM in order to select the most important predictors for assessing the presence/absence of roach in the sampling sites. Before and after variable selection, the SVM were evaluated and compared by two predictive performances namely the percentage of Correctly Classified Instances (CCI %) and Cohen's kappa statistics (k). The obtained results showed that before variable selection, the SVM yielded a reliable performance but the prediction further improved after the combination of SVM with GA. According to the attribute weights, the habitat variables were more responsible than physico?chemical ones in assessing the presence/absence of fish in the streams. GA also presented that roach are more dependent on the habitat variables rather than on water quality ones. Though after variable selection the predictive performances increased, the attribute weights of SVM could be an alternative substitute for GA since all input variables can be evaluated in terms of their weights.