An Error Bound Based on a Worst Likely Assignment

This paper introduces a new PAC transductive error bound for classification. The method uses information from the training examples and inputs of working examples to develop a set of likely assignments to outputs of the working examples. A likely assignment with maximum error determines the bound. The method is very effective for small data sets.

• Type: Journal
• Year: 2008
• Where: JMLR
• Authors: Eric Bax, Augusto Callejas
• Keywords: JMLR 2008, Likely Assignment, PAC Transductive Error, Training Examples