Finding Novel Links in COVID-19 Knowledge Graph Using Graph Embedding Techniques

With the rapid proliferation of scientific literature, it has become increasingly impossible for researchers to keep up with all published papers, especially in the biomedical fields with thousands of citations indexed every day. This has created a demand for algorithms to assist in literature search and discovery. A particular case is the literature related to SARS-CoV-2 where a large volume of papers was generated in a short span. As part of the 2021 Smoky Mountains Data Challenge, a COVID-19 knowledge graph constructed using links between concepts and papers from PubMed, Semantic MEDLINE, and CORD-19, was provided for analysis and knowledge mining. In this paper, we analyze this COVID-19 knowledge graph and implement various algorithms to predict as-yet-undiscovered links between concepts, using methods of embedding concepts in Euclidean space followed by link prediction using machine learning algorithms. Three embedding techniques: the Large-scale Information Network Embedding (LINE), the High-Order Proximity-preserved Embedding (HOPE) and the Structural Deep Network Embedding (SDNE) are implemented in conjunction with three machine learning algorithms (logistic regression, random forests, and feed forward neural-networks). We also implement GraphSAGE, another framework for inductive representation on large graphs. Among the methods, we observed that SDNE in conjunction with feed-forward neural network performed the best with an F1 score of 88.0% followed by GraphSAGE with F1 score of 86.3%. The predicted links are ranked using PageRank product to assess the relative importance of predictions. Finally, we visualize the knowledge graphs and predictions to gain insight into the structure of the graph. © 2022, Springer Nature Switzerland AG.

Drunkenness Face Detection using Graph Neural Networks

The fatalities associated with driving while intoxicated (DWI) are on the rise, leading to a staggering twelve thousand people dying from it and nine lakh people getting arrested every year. DWIs are usually confirmed with the use of breathalyzers, which require the subject to blow into the machine. In light of the current pandemic caused by COVID-19, a susceptible individual may deny blowing into the machine. Thus, the need for a contactless method to detect if someone is drunk arises, so that suspects are prevented from taking advantage of the situation. This also assists law enforcement in the detection of DWI cases. The proposed study is the method to detect intoxication in a given suspect through Graph Neural Networks using facial landmarks. We also present a labeled dataset as a complementary dataset for intoxication detection. This dataset is the first graph-based data available for the detection of alcohol intoxication. Extensive experiments were carried out to validate this approach. © 2021 IEEE.