Customer Relationship Management (CRM) is getting popular, and it is defined as an overall process that relies upon technology to manage the interactions between a business and its customers (Storey 2001), and gain greater insights into its individual customer’s needs (Ryals and Knox 2001). Likewise, Battista and Verhun (2000) mentioned that business leaders had concluded that CRM is a sustainable competitive advantage of business. If business can develop some appropriate offers from the information generated from customer data, customers will be more loyal (Fargo 2000). CRM needs to be supported by techniques. For example, segmentation technique that conditionally divides customers into several groups is a mainstay of CRM (Storey 2001). Unfortunately, K-Means, the most popular clustering method used in cluster-based market segmentation (Green, and Krieger 1995), is non-robust (Garcia-Escudero and Gordaliza 1999), because it is sensitive to initialization (Pena et al. 1999; Dash 2001) and object order (Huang 1998). Segmentation solution might be different in different data conditions (e.g. initialization and/or object order). Thus, K-Means algorithm may converge to non-optimal values (Selim and Ismail 1984). To get better performance in segmenting customers, this thesis investigates the traditional K-Means algorithm and tries to improve it. Therefore, the content of the thesis will be arranged as follows: The motivation and the related literature review is respectively provided in chapter 1 and chapter 2, and a modified K-Means method is proposed in chapter 3. A comparative study between the traditional K-Means and the modified K-Means will be presented in Chapter 4, which indicates that the modified K-Means method is more robust (e.g. the modified K-Means method won’t be greatly influenced by object order and initialization as the traditional K-Means method does), and that modified K-Means can segment data with higher intra-similarity than the traditional K-Means does. A new initialization method, which can be applied to K-Means algorithm, is developed and proposed in chapter 5. Also, Chapter 5 provides a comparative study, indicating that the new initialization method is better than some methods on the performance and the robustness. A discussion and some managerial suggestions about a designed customer data, which is segmented into a few groups by the modified K-Means method, are given in chapter 6. Finally, the conclusion and a discussion of the goals for further research are illustrated in chapter 7.