This paper introduces an online stream mining algorithm based on decision trees using Hoeffding bounds with constant time and space requirements. Overall the paper seems very detailed and the algorithm is introduced including prooves justifying the effectiveness and correctness of the algorihtm. The emperical studies take up a lot of space and are technically sound on first sight. The decision tree learning algorithm builds decision trees recursively at each node. Each incoming sample is used for the current node (or was it level?). How many examples are used at each node (or level) is determined by the Hoeffding bound. Around this algorithm a system called VFDT is build which incorporates the algorithm and enables "very fast descision tree" learning with "constant time and space requirements". If ever going into the direction of stream mining this paper is going to be a reference with already 800+ citations.