Knowledge Transfer for Feature Generation in Document Classification

One important problem in machine learning is how to extract knowledge from prior experience, then transfer and apply this knowledge in new learning tasks. To address this problem, transfer learning leverages information from (supervised) learning on related tasks to facilitate the current learning task. Self-taught learning uses information extracted from (unsupervised) learning on related data. In this paper, we propose a new method for knowledge extraction, transfer and application in classification. We consider document classification where we mine correlation relationships among the words from a set of documents and compile a collection of correlation relationships as prior knowledge. This knowledge is then applied to generate new features for classifying documents in classes/types different from the ones from which we obtain the correlation relationships. Our experiment results show that the correlation-based knowledge transfer helps to reduce classification errors.