Fusing Personal Profile and Group Experience: A Major Recommendation Framework Built on LLM-Generated Data

Major recommendation systems have considered multifaceted user profiles. However, their heavy reliance on questionnaire surveys makes it difficult to extend their application, and also renders recommendation outcomes susceptible to interference from biases. Addressing these issues, this paper proposes a multidimensional decision-fusion major recommendation framework for university applicants. A large language model was innovatively utilized to generate participant responses, constructing a simulated dataset (SimuCareer-25) to overcome the low efficiency and high bias of traditional survey methods. Building upon this, the framework integrates three key decision dimensions: preference, capability, and experience. It then employs fuzzy theory and XGBoost to process multidimensional user features encompassing personality, interests, subject literacy, and skills. Simultaneously, it leverages collaborative filtering to mine group experience. An Analytic Hierarchy Process module is used to perform adjustable weighted fusion of these three dimensions. Through experiments on SimuCareer-25 and real-world data, it was indicated that integrating preference, capability, and historical experience is conducive to generating robust and reliable major recommendations, and that despite structural deviations from real-world samples, the simulated data provides effective feature transferability to support the framework’s practical utility.