PROJECT TITLE :

Hypergraph Learning Methods and Practices

ABSTRACT:

Learning that is carried out on a hypergraph structure can be done through a method known as hypergraph learning. Because of its adaptability and capacity for modeling intricate data correlations, hypergraph learning has garnered an increasing amount of interest in recent years. In this paper, we begin by conducting an in-depth analysis of previously published research on various methods of hypergraph generation, such as those that are distance-based, representation-based, attribute-based, and network-based respectively. Then, we discuss the existing learning methods that can be applied to a hypergraph. These methods include multi-modal hypergraph learning, transductive hypergraph learning, inductive hypergraph learning, and hypergraph structure updating. After that, we present a tensor-based dynamic hypergraph representation and learning framework that has the capability to accurately describe high-order correlation in a hypergraph. We conduct extensive evaluations on a variety of common applications, such as object and action recognition, sentiment prediction on Microblogs, and clustering, in order to investigate the efficacy and efficiency of hypergraph generation and learning methods. In addition to this, we contribute a toolkit for the development of hypergraph learning called THU-HyperG.