Proceedings of the Annual Meeting of the Cognitive Science Society

Numerous classifiers for machine learning are powerful and effectivean important path forward is decreasing the complex-ity and increasing the transparency of the solutions achieved. RELIC (REcursive LInear Classifer) consists of recursivelyapplying a classifier to the training items not successfully accounted for in the previous iteration to find subsets within thetraining data that yield simpler classification schemes. Chooser models are iteratively added and trained on item-to-subsetassignments to learn a mapping between input space and the classifier ensemble. Test examples are passed through the setof choosers to select the appropriate subset-classifier pairing to generate a classification. While applicable to any classifier,we begin by evaluating RELIC using logistic regression and linear SVM to determine whether they perform better underthe recursive approach and become competitive with non-linear classifiers. Application of this approach to non-linearclassifiers and potential implications for the broader science of learning are also addressed.