Education is an essential part of the growth process for students; hence, educational institutions strive to provide high-quality education to students. Nevertheless, accurately assessing students’ performance becomes exceedingly challenging because of the diverse sources and structures of educational data. Moreover, considering the differences in students’ learning abilities, adopting various teaching strategies is necessary. To unearth hidden knowledge from educational data, clustering algorithms have emerged as effective tools. This paper proposes a fuzzy density peak clustering algorithm based on graph distance and natural neighbors (FDPC-GDNN). The core idea is to enhance the separability of clusters via a fuzzy kernel and minimize the influence of outliers. When student portrait assessment and classification are conducted within educational contexts, the categorization of each student is inherently fuzzy, making it difficult to assign them to a specific category. Therefore, the FDPC-GDNN algorithm has significant advantages. We evaluated the FDPC-GDNN on several real-world and synthetic datasets, and the results demonstrated that this method can effectively cluster data of various shapes and densities. Through this approach, educators can gain a better understanding of students’ performance, thereby establishing a solid foundation for instructional decision-making.