Training with Input Selection and Testing (TWIST) Algorithm: A Significant Advance in Pattern Recognition Performance of Machine Learning

摘要: This article shows the efficacy of TWIST, a methodology for design training and testing data subsets extracted from given dataset associated with problem to be solved via ANNs. The we present is embedded in algorithms actualized computer software. Our as implemented software compared current standard methods random cross validation: 10-Fold CV, split into two more advanced T&T. For each strategy, 13 learning machines, representing different families main algorithms, have been trained tested. All were using well-known WEKA package. On one hand falsification test randomly distributed dependent variable has used show how T&T TWIST behaves other strategies: when there no information available on datasets they are equivalent. hand, real Statlog (Heart) dataset, strong difference accuracy experimentally proved. results that superior methods. Pairs similar probability density functions generated, without coding noise, according an optimal strategy extracts most useful pattern classification.