Abstract: The classification algorithms based on association rules generally generate classification association rules by frequent patterns. As mining frequent patterns often suffer from the problem of combinatorial explosion, the efficiency of the algorithms is low. Moreover, the emergence of data streams has posed new challenges for classification algorithms. In contrast to frequent patterns, the number of closed frequent patterns is less, so that the efficiency of algorithms for mining closed frequent patterns is higher. A novel and efficient closed-frequent-patterns based classification algorithm, CBC-DS, is proposed for classifying data streams. The contributions are listed as follows: (1) a single-pass closed frequent itemsets mining process based on reverse lexicographic order FP-tree is introduced for mining classification association rules, which uses a kind of mixed item-ordering searching policy to satisfy the single-pass requirement of data streams and uses the bitmap technology to improve the efficiency; (2) the concept of self-support for filtering rules is proposed to improve the precision. The experimental results show that the bitmap technology can improve the efficiency of the algorithm about twice at least and the average classifying precision can be improved about 0.5% by using self-support. Eventually, the average precision of CBC-DS is about 1% higher than that of CMAR, and CBC-DS is much faster than CMAR.