As an important management method of customer relationship management (CRM), the customer classification is the basis for enterprises to carry out marketing. The classification of customers is conducive to accurate assessment of customer value and facilitate the precise marketing. In this paper, we study the priori structured information hidden in the RFM model dataset, and mark two sets of customer data as a priori category mark, and then get two initial clustering centers. Based on the traditional K-means algorithm, the K value and the initial clustering center are determined with the adaptive method. Combining the two types of constraints of Must-link and Cannot-link, the category markers are transformed into pairs of constraint information. Based on HMRF-KMeans pairs, the constraints and constraint bonuses are introduced to improve the clustering guidance and clustering results. The improved semi-supervised clustering algorithm (RFM-SS-means) was used to test the standard data set, and the Food mart data set was also used to compare the RFM-SS-means algorithm with the traditional K-means algorithm and the two-steps algorithm Class effect. From the experimental results, it can be seen that the CH coefficient of RFM-SS-means is the largest, and the clustering effect is good without prior determination of K value and initial clustering center.