KDD Project Report Using Error-Correcting Codes for Efficient Text Classification with a Large Number of Categories

We investigate the use of Error-Correcting Output Codes (ECOC) for efficient text classification with a large number of categories and propose several extensions which improve the performance of ECOC. ECOC has been shown to perform well for classification tasks, including text classification, but it still remains an under-explored area in ensemble learning algorithms. We explore the use of error-correcting codes that are short (minimizing computational cost) but result in highly accurate classifiers for several real-world text classification problems. Our results also show that ECOC is particularly effective for highprecision classification. In addition, we develop modifications and improvements to make ECOC more accurate, such as intelligently assigning codewords to categories according to their confusability, and learning the decoding (combining the decisions of the individual classifiers) in order to adapt to different datasets. To reduce the need for labeled training data, we develop a framework for ECOC where unlabeled data can be used to improve classification accuracy. This research will impact any area where efficient classification of documents is useful such as web portals, information filtering and routing, especially in open-domain applications where the number of categories is usually very large, and new documents and categories are being constantly added, and the system needs to be very efficient.