Graph Neural Networks for Soft Semi-Supervised Learning on Hypergraphs
| Autor: | Shahab Asoodeh, Anand Louis, Tingran Gao, Partha Pratim Talukdar, Naganand Yadati |
|---|---|
| Rok vydání: | 2021 |
| Předmět: |
Hypergraph
Theoretical computer science Computer science 02 engineering and technology Semi-supervised learning 010501 environmental sciences 01 natural sciences Core (graph theory) 0202 electrical engineering electronic engineering information engineering Code (cryptography) Key (cryptography) Graph (abstract data type) Probability distribution 020201 artificial intelligence & image processing Pairwise comparison MathematicsofComputing_DISCRETEMATHEMATICS 0105 earth and related environmental sciences |
| Zdroj: | Advances in Knowledge Discovery and Data Mining ISBN: 9783030757618 PAKDD (1) |
| DOI: | 10.1007/978-3-030-75762-5_36 |
| Popis: | Graph-based semi-supervised learning (SSL) assigns labels to initially unlabelled vertices in a graph. Graph neural networks (GNNs), esp. graph convolutional networks (GCNs), are at the core of the current-state-of-the art models for graph-based SSL problems. GCNs have recently been extended to undirected hypergraphs in which relationships go beyond pairwise associations. There is a need to extend GCNs to directed hypergraphs which represent more expressively many real-world data sets such as co-authorship networks and recommendation networks. Furthermore, labels of interest in these applications are most naturally represented by probability distributions. Motivated by these needs, in this paper, we propose a novel GNN-based method for directed hypergraphs, called Directed Hypergraph Network (DHN) for semi-supervised learning of probability distributions (Soft SSL). A key contribution of this paper is to establish generalisation error bounds for GNN-based soft SSL. In fact, our theoretical analysis is quite general and has straightforward applicability to DHN as well as to existing hypergraph methods. We demonstrate the effectiveness of our method through detailed experimentation on real-world datasets. We have made the code available. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|