Online shopping sites are faced with a significant problem: When offering experience products, i.e., products that lack a helpful description in terms of easily accessible factual properties (e.g., wine, cigars, and movies), a lot of work and time needs to be invested to provide such information. Two very popular approaches are the introduction of sophisticated categorization systems (e.g., fruity, woody, and peppery for wines) along with manual product classification performed by experts and the addition of user feedback mechanisms (e.g., ratings or textual reviews). While user feedback typically is easy to collect, for purposes of product search, it cannot be used as easily as this is possible with a systematic categorization scheme. In this paper, we propose an effective method to automatically derive product classifications of high quality from many different kinds of user feedback. Our semi-supervised method combines advanced data extraction methods with state-of-the-art classification algorithms and only requires a small number of training examples to be created manually by experts. We prove the benefits of our approach by performing an extensive evaluation in the movie domain.