We propose a stream mining method that learns opinionated product features from a stream of reviews. Monitoring the attitude of customers towards products is a field of much interest, but the products themselves may come in and out of the market. We rather investigate which (implicit) features of the products are important for the customers, and monitor how customer attitude towards such features evolves. To this purpose, we use a two-level stream clustering algorithm that extracts features and subfeatures from an opinionated stream, and couple it with dedicated feature-specific classifiers that assess the polarity of each extracted (sub)feature. We evaluate our method on a stream of reviews and we elaborate on how changes in the arrival rate of features (drift) affects algorithm performance.