Automated weak signal detection and prediction using keyword network clustering and graph convolutional network

Weak signals are rarely identified in the initial stage of growth and appear significant over time, unlike strong signals clearly observed in past trends. Weak signals are important cues that need to be analyzed to rapidly and accurately predict changes in the uncertain future. Researchers have developed various methods for identifying cues that can be significantly used for prediction. However, in many cases, they heavily depend on the opinions of experts or are applicable only to weak signals in specific fields. This study proposes a weak signal detection method that extracts weak signals by selecting significant keywords from literature database and grouping relevant keywords. Furthermore, this study presents a weak signal prediction method for predicting the growth of specific weak signals by investigating and learning the growth of the extracted weak signals over 10 years. To verify the proposed method, we extracted weak signals for 10 years (2001-2010) from SCOPUS publication data from 1996 to 2009 and applied machine learning using a graph convolutional network (GCN) model with the growth data of the extracted weak signals. The results showed that the proposed methods can effectively detect and predict weak signals.