Comparative Analysis of Vectorization Methods for Academic Supervisor Recommendations

Abstract

Selecting final project supervisors often poses challenges for students due to limited lecturer quotas and difficulties in finding suitable expertise matches. This study proposes using the Cosine Similarity method with vectorization approaches such as Bidirectional Encoder Representations from Transformers (BERT), FastText, Bag of Words (BoW), Term Frequency-Inverse Document Frequency (TF-IDF), and Word2Vec to enhance the accuracy of recommendation systems. Data sourced from Google Scholar underwent scraping, preprocessing, and vectorization to evaluate the most effective method for understanding context and recommending relevant supervisors. The analysis revealed that BERT and Word2Vec based approaches achieved superior performance, delivering a perfect hit ratio (1.00) and overcoming the limitations of TF-IDF and BoW in capturing technical language. This recommendation system is expected to streamline the supervisor selection process, minimize mismatches, and effectively support academic advisory processes across educational institutions.