A Hypergraph-Based Alternative to Convolutional Layer in Image Classification Models

Hypergraph modeling has emerged as an effective approach for representing data by capturing higher-order relationships among multiple data instances. Recent methods utilizing hypergraphs demonstrate applications in signal processing such as image denoising, compression and spectral clustering. To perform learning tasks on hypergraph representation of image data, hypergraph neural networks have been introduced to enhance performance. This paper explores the effectiveness of hypergraph signal representations as input features for neural network-based image classification models. The results show that a mixed hypergraph network achieves performance comparable to traditional convolutional neural networks, while requiring less computation, as evidenced by shorter training time.