Extraction of Representative Keywords Considering Co-occurrence in Positive Documents

In linear text classification, user feedback is usually used to tune up the representative keywords (RK) for a certain class. Despite some algorithms (e.g. Rocchio) deal well with user positive and negative feedback to adjust the RKs, few researches have investigated how to adjust RKs only based on a small positive responses which is a popular case in the real-world application (e.g. users tend to click their interested URL). In this work, we describe a method of extracting representative keywords for a user from a small set of his positive feedback documents. Experiments on the Reuters-21578 collection illustrate that our approach is better than other two famous methods (Rocchio and Widrow-Hoff) with 24.8% and 14.5% improvement, respectively.