Managing Training Examples for Fast Learning of Classifiers Ranks
Paper deals with the problem of learning ranks of classifiers in ensembles. The problem of ordering of objects to classify is discussed. Two marginal approaches for learning, batch and incremental, with corresponding ordering strategies are analyzed. Presented algorithm lays between marginal methods, and it orders training examples by the deviation of classifiers opinions to match restrictions on learning time, cost and quality. Few aspects of this algorithm are experimentally investigated: classifiers ranks after learning, learning quality, ensemble accuracy and dependence between rank recalculation budget and ensemble accuracy. It was found, that descending order of examples provides fast rank learning with the best learning quality.
Published in 1999.