In recent years, with the continuous development and growth of the Non-Fungible Token market, it is of great significance to conduct clustering research on its users in order to improve marketing efficiency and enhance customer service. In the field of user clustering research, the classic unsupervised algorithm K-Means has been widely used. However, the K-Means algorithm has some deficiencies, such as need to pre-select the value of K, lack utilization of prior knowledge, limited interpretability of clustering results. Therefore, this paper proposes an improved K-Means algorithm named PKK-Means. Firstly, the algorithm uses the key users identified by prior knowledge to analyze the distribution of key users and other users in each feature item, thereby obtaining the initial cluster centers and the initial number of K. Then, based on the attribute items of the user sample distribution matrix, the algorithm calculates the sum of squared errors within clusters after K-partitioning, and adjusts the similarity measurement of users accordingly. Finally, this paper improves the user value measurement model RFM named RFMCO by introducing the digital currency market value and the on-chain index. Experiments show that compared to traditional algorithms, the K-Means algorithm based on prior knowledge does not require the pre-selection of K value and achieves significant improvement in clustering effectiveness, especially when the dataset is large, the interpretability of the algorithm is significantly improved.