<![CDATA[The increasing volume of academic documents in University of Udayana Digital Library (e-Perpus UNUD) has increased the difficulty for students to efficiently identify relevant prior research aligned with their research topics. Conventional keyword-based search mechanisms often fail to capture contextual relevance, resulting in information overload and suboptimal retrieval quality. To address this challenge, this study presents the design and implementation of a thesis recommendation system based on a modified BM25 algorithm called Position-Weighted Retrieval BM25 (PoWeR-BM25). The proposed model introduces an adaptive positional weighting scheme that emphasizes query term occurrences appearing at the beginning and end of a document, where essential contextual information is typically concentrated. The system was developed using an information retrieval framework and evaluated across multiple configurations combining different data representations (title, abstract, and title–abstract) and preprocessing techniques (stemming and non-stemming). Experimental results demonstrate that POWER-BM25 consistently outperforms traditional BM25 and TF-IDF in terms of Mean Average Precision (MAP) and F1-Score, particularly when applied to stemmed abstract data. The best performing configuration was subsequently deployed as recommendation feature integrated into University of Udayana Digital Library via a Web Service API, enabling users to obtain more accurate and contextually relevant thesis recommendations. These findings highlight the practical importance of incorporating positional term weighting into probabilistic retrieval models as a lightweight yet effective approach to improving academic recommendation systems without relying on computationally intensive machine learning models.]]>
