Hypergraph Representation Learning from Noisy Node Attributes

Qian-xi Tang, Chun-Yang Zhang

Abstract: In recent years, hypergraph representation learning (HGRL) has become a focus of academic research, which aims to extract high-order topological patterns and attributes from hypergraph into low-dimension representation vectors. However, most existing methods ignore the uncertainties in the hypergraph data, thus failing to effectively leverage attribution features hidden in hypergraphs. For example, in a citation hypergraph, there exist uncertain semantics in node attributes regarding to papers. Therefore, we propose a new Fuzzy HGRL model, called FAHGN, which introduces fuzzy logic to grasp the attribute uncertainties. Specifically, the proposed FAHGN fuzzifies node attributes of the hypergraph as fuzzy input hypergraph signals, and makes full use of a spectral graph convolution operator to aggregate the fuzzy input signals to generate node representations. Through this way, it sufficiently takes into account the feature-level uncertainties in hypergraphs and provides more expressive representations for effectively achieving downstream tasks. The experimental results on five real-world datasets validate the effectiveness of the proposed FAHGN against competitive baseline models.

Submission Number: 73