Methods and systems for cost-sensitive boosting

A method for multi-class, cost-sensitive learning based on iterative example weighting schemes applied to a chosen data set, comprising:

• a) obtaining an expanded data set, which is defined by enhancing each example in an original data set with as many data points as there are possible labels for any single instance;
• b) repeatedly calling a component importance-weighted classification learning algorithm with the expanded data set, with example weights, wherein each non-optimally labeled example is given a weight which equals a half times an original misclassification cost for said labeled example times p−1 norm of average prediction of current hypotheses, and each optimally labeled example is given a weight which equals sum of weights for all the non-optimally labeled examples for the same instance;
• and
• c) outputting all classifier representations obtained through iterations of step b), each of said classifier representations being a representation of classifier and together representing an average classifier.

A program storage memory device readable by a machine, tangibly embodying a program of instructions executable by the machine to perform a method for multi-class, cost-sensitive learning based on iterative example weighting schemes applied to a chosen data set, comprising:

• a) obtaining an expanded data set, which is defined by enhancing each example in an original data set with as many data points as there are possible labels for any single instance;
• b) repeatedly selecting one or more sub-samples from the expanded data set using weighted sampling according to a predetermined example weighting scheme, wherein each non-optimally labeled example is given a weight which equals a half times an original misclassification cost for said labeled example times p−1 norm of average prediction of current hypotheses, and each optimally labeled example is given a weight which equals sum of weights for all the non-optimally labeled examples for the same instance;
• c) executing a component classification learning algorithm on the sub-sample obtained in step b) and obtaining a hypothesis representing a classifier;
• and
• d) outputting all classifier representations obtained in iterations of steps b) and c), each of said classifier representations being a representation of classifier.