Abstract

One of the main challenges faced by new researchers is immersing themselves in the existing literature relevant to their field of interest. The vastness and continuous growth of knowledge in their field can be overwhelming, making it difficult to identify the most pertinent research papers within their research themes. To address this issue, research paper recommender systems have emerged as valuable tools. These systems allow researchers to find relevant papers based on their specific interests or research themes by analyzing various aspects such as titles, abstracts, and full texts. The quality of the dataset used is crucial for the development, testing, and refinement of these systems to ensure optimal results. Dataset quality directly impacts the accuracy and reliability of a recommender system. In this thesis, I propose a novel approach for constructing datasets using citation graph networks. These networks consist of nodes representing research papers and edges representing citations between them. By leveraging citation graph networks, we gain a more comprehensive understanding of the relationships and influences among different papers compared to traditional methods that rely solely on keyword searches. To evaluate the effectiveness of the citation graph network method, I compared it with the traditional keyword search approach for dataset construction. Additionally, I assessed the effectiveness of three recommender system algorithms: user-based collaborative filtering, combined with PageRank and personalized PageRank algorithms. The experimental findings provide clear evidence that utilizing citation graph network datasets significantly enhances the efficacy of research paper recommender systems. This improvement simplifies the process of finding relevant literature for researchers, potentially accelerating scientific discovery.