In this work, we propose a new method for extracting user preferences from a few documents that might interest users. For this end, we first extract candidate terms and choose a number of terms called initial representative keywords (IRKs) from them through fuzzy inference. Then, by expanding IRKs and reweighting them using term co-occurrence similarity, the final representative keywords are extracted. Performance of our approach is heavily influenced by effectiveness of selection method for IRKs so we choose fuzzy inference because it is more effective in handling the uncertainty inherent in selecting representative keywords of documents. The problem addressed in this paper can be viewed as the one of finding a representative vector of documents in the linear text classification literature. So, to show the usefulness of our approach, we compare it with two famous methods Rocchio and Widrow-Hoff - on the Reuters-21578 collection. The results show that our approach outperforms the othe...