Most previous works on text categorization applied term occurrence frequency and inverse document frequency for representing imporlanee of terms. This paper presents an analysis of inverse class frequency in centroid-based text categorization. Two aims of this paper are concerned. The fint one is to find appropriale functions of inverse class frequency. The other is to tind the key facton for using invene class frequency. The experimental results show that the key factors, which improve classification accuracy, are the numben of fewclass terms and most-class terms. When large numhen of fewclass terms and most-class terms are obtained, logarithmic hnrlion of inverse class frequency is the most emective when it is combined with term frequency. The square root of inverse class frequency incorporated into TFIDF, works well in the case that dab sets include a smaU number of few-class terms and mostclass terms. To increase lhe numbers of these endive terms, some methods are involved i.e. using higher gram models, small number of classes and large number oftraining set size.