Audience segmentation is often applied on Web portals to gain insights into the audience, support targeted marketing and in general provide on-line recommendations to the users. We propose an approach to audience segmentation that is based on using machine learning on topic profiles of the visited content. Our preliminary experiments on a small sample of log-data show that the proposed approach is promising and the proposed combination of features capturing short term and long-term user interest gives better results than using only the short-term interests of the user.