Different from the group formation approaches led by teachers, learners' generative learning objectives and the independent choice of collaborative partners are important in the Internet learning environment. This study takes the cMOOC (connectivist massive open online course) 5.0 "Internet + education: dialogue between theory and practice" as a case course. Data on learners' interactive behaviours, generative texts, and personal attributes were collected through the course platform and questionnaires. Social network analysis and natural language processing were adopted to build a variety of collaborator relationship networks, and the quadratic assignment procedure was used to reveal the impact of the social, topic and attribute relationships of learners before collaboration on the collaborative relationship. The results show that individuals with close social relations, similar topics, similar age and learning experiences before collaboration are more likely to establish collaborative relationships; among them, close social relations and close ages are the key driving factors affecting the choice of partners. In addition, topic similarity can further promote the establishment of collaborative relationships among individuals with close social relations. The findings will help community-based courses optimise partner recommendation strategies and promote collaborative problem-solving and collective knowledge creation.