# Deep Hypergraph Structure Learning

**Authors:** Zizhao Zhang, Yifan Feng, Shihui Ying, Yue Gao  
**arXiv:** [2208.12547](https://arxiv.org/abs/2208.12547)  
**Date:** 2022-08-26  
**Summary page:** https://www.summarizepaper.com/en/paper/2208.12547v1/

## Key points

- High-order correlation is superior in data representation learning
- Hypergraphs are widely used in recent decades for this purpose
- Performance of hypergraph-based methods heavily relies on the quality of the hypergraph structure
- Generating a suitable hypergraph structure among data is challenging due to missing and noisy data that may lead to "bad connections"
- Zizhao Zhang et al. propose a general paradigm for deep hypergraph structure learning called DeepHGSL that optimizes the hypergraph structure for hypergraph-based representation learning
- The HyperGraph Information Bottleneck (HIB) principle guides the optimization of the hypergraph structure by constructing a loss function that minimizes noisy information in it
- DeepHGSL enhances correct connections and weakens wrong connections during training phase, thus optimizing and improving upon an otherwise heavily noisy structure
- The authors evaluate their model on four benchmark datasets for representation learning and demonstrate its effectiveness and robustness on both graph- and hypergraph-structured data compared with other state-of-the-art methods.
- DeepHGSL provides a promising solution to generate suitable high-quality structures for effective representation learning using deep neural networks while addressing challenges posed by missing or noisy data through HIB-guided optimization of these structures.

## Summary

The use of high-order correlation has proven to be superior in data representation learning, with hypergraphs being widely used in recent decades. However, the performance of hypergraph-based representation learning methods, such as hypergraph neural networks, heavily relies on the quality of the hypergraph structure. Generating a suitable hypergraph structure among data is still a challenging task due to missing and noisy data that may lead to "bad connections" in the hypergraph structure and disrupt the learning process. Therefore, it becomes crucial to reveal the high-order structure or the hypergraph behind observed data. To address this issue, Zizhao Zhang et al. propose a general paradigm for deep hypergraph structure learning called DeepHGSL that optimizes the hypergraph structure for hypergraph-based representation learning. They extend the information bottleneck principle for robustness issues to the hypergraph case and name it as the HyperGraph Information Bottleneck (HIB) principle. This principle guides the optimization of the hypergraph structure by constructing a loss function that minimizes noisy information in it. The proposed method enhances correct connections and weakens wrong connections during training phase, thus optimizing and improving upon an otherwise heavily noisy structure. The authors evaluate their model on four benchmark datasets for representation learning and demonstrate its effectiveness and robustness on both graph- and hypergraph-structured data compared with other state-of-the-art methods. Furthermore, DeepHGSL provides a promising solution to generate suitable high-quality structures for effective representation learning using deep neural networks while addressing challenges posed by missing or noisy data through HIB-guided optimization of these structures. Overall, DeepHGSL offers an efficient approach to optimize a suitable high quality graph structure for better performance in representation learning tasks involving large amounts of missing or noisy data.