<![CDATA[Choosing a final project topic is an early decision that shapes a student's research direction. This study builds and evaluates a prototype recommender for undergraduate Informatics final project topics based on content-based filtering, using each student's interests and course track record as the query. Four methods are compared, namely TF-IDF with cosine similarity, BM25, Latent Semantic Analysis (LSA), and a hybrid of BM25 and LSA, together with a random baseline. The offline evaluation uses 20 synthetic corpora, each with 400 topics from eight Informatics areas and 150 student profiles, and reports Precision@k, MRR, and NDCG@10. Because real data were unavailable and Sentence-BERT could not be downloaded in the research environment, this is a simulation study and does not evaluate Sentence-BERT; its evaluation script is provided. BM25 obtained the highest Precision@5 (0.927), followed by the hybrid (0.914), TF-IDF (0.895), LSA (0.876), and random (0.209); all pairwise differences in Precision@5 remained significant after Holm correction. The hypothesis that latent semantic methods are more robust to term variation was not supported, because LSA declined the most as term variation increased. These findings are preliminary and bounded by the assumptions of the data generator; validation on real final project repositories with expert relevance judgments is still required.]]>
Copyrights © 2026
