Comparative Analysis of Research Article Matching using SIF, RNN, Attention, and Hybrid Methods

Search engines make it easier to conduct literature reviews. However, for niche topics, search results are often poor. Snowballing can help, but it is limited by the initial articles, especially by the authors' access when they were written. As an alternative, research paper databases have provided recommendation features; however, these are limited to their own articles. A tool to search for similar articles without relying on a specific database would be helpful, but before that, a proper method to match similar articles must be found. This research aims to match similar articles based on title, authors, and keywords using deep learning methods, which are SIF, RNN, Attention, and Hybrid methods, and evaluate them. This study also compares the combinations of features used in matching. The attention method using only the article title as a feature yielded the best result. The attention method was also faster than the hybrid method for training and use. Using only one feature should be even faster. In addition, the title field was found to be the best feature for predicting similarity matches. The author name feature was bad on its own but could improve the results when combined with the title. The keyword feature was found to be almost as good as the title, but combining them did not result in significant improvement.