Reproduce “Hyperbolic Interaction Model For Hierarchical Multi-Label Classification” on OpenResearch

About this paper

Boli Chen, Xin Huang, Lin Xiao, Zixin Cai, Liping Jing

Researchers at Beijing Jiaotong University developed HyperIM, a hyperbolic interaction model for hierarchical multi-label classification that embeds labels and words in Poincaré ball. This approach consistently achieved superior performance across three benchmark datasets by effectively capturing hierarchical relationships and fine-grained semantic connections.

Key insights

• Hyperbolic space, with its constant negative curvature, is inherently more suitable than Euclidean space for embedding and preserving hierarchical structures with significantly less distortion.
• Jointly embedding words and labels in hyperbolic space, and measuring their similarity via geodesic distance, enables effective capture of fine-grained semantic connections and the creation of highly label-aware document representations.
• Explicitly incorporating and initializing both word and label hierarchies (e.g., from WordNet) during the embedding process in hyperbolic space provides a substantial performance boost beyond merely using the hyperbolic geometry itself.

Summary by alphaXiv.