The present study sought to reveal the relationship between two types of narcissism (grandiose and hypervigilant) and the positive and negative wording items (PSE and NSE) of Rosenberg’s self-esteem scale (RSES). We conducted a series of clustering analyses using three datasets with a k-means++ clustering method. Each dataset included both RSES and one or two narcissism scales: (a) RSES and two types of narcissism scales (N = 900), (b) RSES and a grandiose narcissism scale (N = 400), and (c) RSES and a hypervigilant narcissism scale (N = 600). Based on the result of a preliminary cluster analysis performed on a RSES dataset extracted both from the three datasets mentioned above and other data sets (N = 5,337), we determined the number of clusters as five. The results of a cluster analysis of dataset (a) revealed that individuals who were in the “PSE-predominant self-evaluation cluster” (i.e., individuals who had high self-evaluation on the PSE items and low self-evaluation on the NSE items) exhibited a greater tendency for both grandiose and hypervigilant narcissism compared with those in the other four clusters. In contrast, individuals in the “NSE-predominant self-evaluation cluster” exhibited a reduced tendency for both grandiose and hypervigilant narcissism compared with those in the other clusters. Analyses of the other two datasets (b and c) revealed similar clusters to those in dataset (a). These results suggest that the relationship between self-esteem and narcissism can be better interpreted using a two-factor solution for RSES compared with using an average score of all items of the RSES.