Publication Date

Spring 5-25-2021

Degree Type

Master's Project

Degree Name

Master of Science in Computer Science (MSCS)

Department

Computer Science

First Advisor

Katerina Potika

Second Advisor

Chris Pollett

Third Advisor

William Andreopoulos

Keywords

higher order link prediction, triangle embeddings, GloVeNoR, Simplex2Vec

Abstract

Social media, academia collaborations, e-commerce websites, biological structures, and other real-world networks are modeled as graphs to represent their entities and relationships in an abstract way. Such graphs are becoming more complex and informative, and by analyzing them we can solve various problems and find hidden insights. Some applications include predicting relationships and potential links between nodes, classifying nodes, and finding the most influential nodes in the graph, etc.

A large amount of research is being done in the field of predicting links between two nodes. However, predicting a future relationship among three or more nodes in a graph is a more recent active research topic. Relationships that involve more than two nodes is called a higher-order link. One of the approaches, that we follow in this work, is that of mapping the graph entities, such as nodes, edges, and triangles, into a low dimensional space by generating embeddings vectors. In that way, we work with vectors and reduce the higher-order link prediction to a classification problem.

The primary objective of this project is to utilize the GloVeNoR node embedding technique, as well as Simplex2Vec triangle embedding technique, to perform higher- order link prediction, i.e., to predict the possibility of interaction. Additionally, we evaluate the predictions generated by our methods and compare them with existing higher-order link prediction approaches using benchmark datasets. Based on our experiments, we show that the triangle embeddings generated using the techniques discussed in the report increase the average performance over the five datasets evaluated using the AUC-PR relative to random baseline as a metric for higher-order link prediction by 48%.

Share

COinS