Application of Stacked Generalization to a Protein Localization Prediction Task

A stacked generalization data mining approach was applied to a simplified version of the KDD Cup 2001 protein localization task. Four level-0 models were developed: an Artificial Neural Network, a Decision Tree, a Nearest Neighbor Classifier, and a Hybrid model that trained an Artificial Neural Network using those inputs selected as important by the Decision Tree. These models were developed from both a random sampling of the training dataset and a sampling designed to ensure an equal distribution of the localization variable. The predictions of these level-0 models were used to develop three level-1 generalizers: an Artificial Neural Network, a Decision Tree, and a Naïve Bayesian Classifier. The accuracy rates of the various models and generalizers suggest a modestly improved performance when using stacked generalization. Among the level-0 models, the Nearest Neighbor and the Hybrid performed slightly better than the Artificial Neural Network and Decision Tree. The performance of the three level-1 generalizers was roughly equivalent. Overall, both level-0 models and level-1 generalizers achieved better accuracy when equal distribution sampling was not performed.