The School of Computer Science is pleased to present…
Date: Tuesday, May 6th, 2025
Time: 11:30 AM
Location: Essex Hall, Room 122
The enhanced performance of Graph Neural Networks (GNNs) has attracted significant attention in various research fields. However, in many real-world applications, the inherent relationships among objects in higher dimensions are typically not captured by standard graphs since edges connect only two vertices. Hypergraphs, instead, tackle this limitation by introducing a hyperedge that can connect an arbitrary number of nodes. This raises a key question about the information loss caused by the limitations of graph-based representations. The similarity between the definitions of subgraphs and hyperedges inspired us to introduce the Densest Overlapping Subgraphs (DOS) as a primary framework used to convert a graph into a hypergraph.
In this work, we propose to address this problem by proposing an information retrieval framework designed to transform graph structures into hypergraph representations. This framework leverages the concept of DOS as a key method for constructing hyperedges, thereby taking advantage of the richer representational capacity of hypergraphs, yielding enhanced graph machine learning tasks. Preliminary results on synthetic and real-world datasets demonstrate that our approach consistently yields improved performance over traditional graph-based methods while offering flexible control over subgraph size, density, and overlap.
Internal Reader: Dr. Dan Wu
External Reader: Dr. Ning Zhang
Advisor: Dr. Luis Rueda
Chair: Dr. Muhammad Asaduzzaman