Utilising Filter Inferred Information in Nature-inspired Hybrid Feature Selection

The development of modern smart systems such as smart grid, medical diagnosis assessment tools or quality control systems in manufacturing relies heavily on data and knowledge attained from, possibly, large amounts of information. Several well known but hard to solve problems often arise during this process, a prominent one being the problem of high dimensionality. Although different methods of feature selection are both proposed and employed in order to ameliorate its effects, it remains an open problem. Recently, hybrid procedures combining both filters as a preprocessing step and a wrapper as a refining step have proven to be an effective approach. Along the problem of filter and wrapper selection, the problem of incorporating knowledge inferred by the filter into the wrapper is an interesting one. In this paper, a novel approach is proposed that relies on filter information in order to generate the initial population of a nature-inspired algorithm utilised as the wrapper. Promising results were obtained on several real-world datasets that demonstrate the effectiveness of the proposed approach both in terms of classification quality and dimensionality reduction. Results of the statistical analysis of performance when compared with other approaches further emphasize the observed benefits.