Publication Date

Spring 2020

Degree Type

Master's Project

Degree Name

Master of Science (MS)

Department

Computer Science

First Advisor

Katerina Potika

Second Advisor

Chris Pollett

Third Advisor

William Andreopoulos

Keywords

link prediction, triangle prediction, node2vec, graph2vec, graph neural networks

Abstract

Link prediction is an emerging field that predicts if two nodes in a network are likely to be connected or not in the near future. Networks model real-world systems using pairwise interactions of nodes. However, many of these interactions may involve more than two nodes or entities simultaneously. For example, social interactions often occur in groups of people, research collaborations are among more than two authors, and biological networks describe interactions of a group of proteins. An interaction that consists of more than two entities is called a higher-order structure. Predicting the occurrence of such higher-order structures helps us solve problems on various disciplines, such as social network analysis, drug combinations research, and news topic connections. Moreover, we can use our methods to get more knowledge about news topics during the COVID-19 pandemic.

Higher-order link prediction can be accomplished using neural networks and other machine learning techniques. The primary focus of this project is to explore repre- sentations of three-node interactions, called triangles (a special case of higher-order structure). We propose new methods to embed triangles: by generalizing node2vec algorithm using different operators to learn an embedding for a triangle, and by using 1-hop subgraphs of the triangles to learn embeddings using graph2vec algorithm and graph neural networks. The performance of these techniques is evaluated against the benchmark scores on various datasets used in the bibliography. From the results, it is observed that the node2vec based triangle embedding algorithm performs better or similar on most of the datasets compared to benchmark models.

Download

Share

COinS