Team role research has been a hot research issue in recent years, and a good team background is very important for team role research. A good team is not a simple collection of multiple people but many people with different roles and functions (leadership, group leader, staff), they have close cooperation and strict relationship between superiors and subordinates. Team role research helps us to understand the function and behavior of the members of the team. Compared with the traditional walking method, Mjwalk can walk on multiple independent team networks, showing advantages in multinetwork joint learning. The conv_att method proposed in this paper solves the downstream role recognition task and improves the recognition effect. We evaluate the performance of the Mjwalk and conv_att framework with real team data, and improve f1_score and accuracy.