Prediction and importance of predictors in approaches based on computational intelligence and machine learning

Resumo

Machine learning and computational intelligence are rapidly emerging in plant breeding, allowing the exploration of big data concepts and predicting the importance of predictors. In this context, the main challenges are how to analyze datasets and extract new knowledge at all levels of research.  Predicting the importance of variables in genetic improvement programs allows for faster progress, carrying out an extensive phenotypic evaluation of the germplasm, and selecting and predicting traits that present low heritability and/or measurement difficulties. Although, simultaneous evaluation of traits provides a wide variety of information, identifying which predictor variable is most important is a challenge for the breeder. The traditional approach to variable selection is based on multiple linear regression. It evaluates the relationship between a response variable and two or more independent variables.  However, this approach has limitations regarding its ability to analyze high-dimensional data and not capture complex and multivariate relationships between traits. In summary, machine learning and computational intelligence approaches allow inferences about complex interactions in plant breeding. Given this, a systematic review to disentangle machine learning and computational intelligence approaches is relevant to breeders and was considered in this review. We present the main steps for developing each strategy (from data selection to evaluating classification/prediction models and quantifying the best predictor).