Academic article recommender system

With advances in information technologies and scientific developments, thousands of publications are published by researchers every year. Due to the exponentially increasing amount of information, accessing relevant academic documents online makes it difficult and time-consuming. Recommender systems are widely used in solving such problems. Recommender systems are developed to present documents containing the necessary information in line with the preference of researchers. These systems consist of algorithms that filter utility content from the information in large data pools and find the most meaningful and relevant product for the user. With these algorithms, the after movements of users are predicted by considering information such as historical information, interests, and click data. In the academic article recommendation systems, related articles are suggested by using the information found in the researchers' profiles. The use of these systems enables researchers to access publications quickly and effectively. In this study, a new article recommendation system consisting of hybrid combinations of content-based methods is designed. In the proposed system, metadata of the articles collected from the ARXIV dataset was used. In addition to word embedding algorithms such as Word2vec, Term Frequency – Inverse Document Frequency (TF-IDF), Doc2vec, and Global Vectors (GloVe), FastText, subject modeling techniques such as Latent Dirichlet Allocation (LDA) and Non – Negative Matrix Factorization (NMF) were also applied. The hybrid methods proposed in the study were compared with the other algorithms used for an in-depth analysis. In evaluating the proposed system, user profiles created within the scope of the study were used. In this way, it is ensured that the relevant suggestions are listed for a new user or researcher. Offline experiments were conducted in the study. In experiments where Word2vec and LDA methods and TF-IDF and LDA methods were combined as a hybrid, better performance values were obtained than other methods. The numerical values obtained from the experiments carried out within the scope of the study confirm that the proposed system is promising when compared with the literature.