Accurate Matching and Recommendation for University Innovation Projects: A Multimodal Mobile Learning Approach

University innovation projects are increasingly complex and require personalized recommendation systems that align students’ academic strengths with appropriate project opportunities. Existing recommendation methods typically rely on limited academic criteria, which fail to capture the full diversity of student learning characteristics, resulting in suboptimal matching. To address this, the research proposes an integrated mobile learning-based system that utilizes structured academic data from students’ marksheets. The framework distinguishes multiple modalities, including overall academic performance (GPA, weighted GPA), subjectdomain proficiency (subject-wise scores), and temporal learning trends (semester-wise performance variations). Although these features originate from a single dataset, they represent distinct perspectives on student learning. Each modality is transformed into a feature representation, and these heterogeneous representations are combined into a unified multimodal student profile through vector-level fusion. A collaborative filteringbased recommendation engine, powered by cosine similarity, then generates personalized project suggestions. The system’s effectiveness is evaluated using standard recommender metrics such as Precision@5, Recall@5, F1@5, NDCG@5, MAP@5, hit rate, and coverage. Experimental results demonstrate the proposed method outperforms baseline approaches, achieving improvements of 12.6% in Precision@5 and 9.8% in NDCG@5, indicating enhanced ranking accuracy and recommendation relevance. Overall, the findings confirm that multimodal academic representations significantly improve personalized and large-scale innovation project recommendations in higher education.