Abstract. One of the main problems that modern e-mail systems face is the management of the high degree of spam or junk mail they recieve.Those systems are expected to be able to distinguish between legitimate mail and spam; in order to present the nal user as much interesting information as possible. This study presents a novel hybrid intelligent system using both unsupervised and supervised learning that can be easily adapted to be used in an individual or collaborative system. The system divides the spam ltering problem into two stages: rstly it divides the input data space into di erent similar parts. Then it generates several simple classi ers that are used to classify correctly messages that are contained in one of the parts previously determined. That way the e ciency of each classi er increases, as they can specialize in separate the spam from certain types of related messages. The hybrid system presented has been tested with a real e-mail data base and a comparison of its results with those obtained from other common classi cation methods is also included. This novel hybrid technique proves to be e ective in the problem under study.