Scientific Impact of Graph-Based Approaches in Deep Learning Studies -- A Bibliometric Comparison

Applying graph-based approaches in deep learning receives more attention over time. This study presents statistical analysis on the use of graph-based approaches in deep learning and examines the scientific impact of the related articles. Processing the data obtained from the Web of Science database, metrics such as the type of the articles, funding availability, indexing type, annual average number of citations and the number of access were analyzed to quantitatively reveal the effects on the scientific audience. It's outlined that deep learning-based studies gained momentum after year 2013, and the rate of graph-based approaches in all deep learning studies increased linearly from 1% to 4% within the following 10 years. Conference publications scanned in the Conference Proceeding Citation Index (CPCI) on the graph-based approaches receive significantly more citations. The citation counts of the SCI-Expanded and Emerging SCI indexed publications of the two streams are close to each other. While the citation performances of the supported and unsupported publications of the two sides were similar, pure deep learning studies received more citations on the journal publication side and graph-based approaches received more citations on the conference side. Despite their similar performance in recent years, graph-based studies show twice more citation performance as they get older, compared to traditional approaches. Annual average citation performance per article for all deep learning studies is 11.051 in 2014, while it is 22.483 for graph-based studies. Also, despite receiving 16% more access, graph-based papers get almost the same overall citation over time with the pure counterpart. This is an indication that graph-based approaches need a greater bunch of attention to follow, while pure deep learning counterpart is relatively simpler to get inside.