Shadow Cones: A Generalized Framework for Partial Order Embeddings

Authors: Tao Yu, Toni J.B. Liu, Albert Tseng, Christopher De Sa

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments on datasets of various sizes and hierarchical structures show that shadow cones consistently and significantly outperform existing entailment cone constructions. and Achieve state-of-the-art results on a wide range of graph embedding tasks, lending both empirical and theoretical support to the advantages of all four embedding schemes. and 5 EXPERIMENTS This section showcases shadow cones ability to represent and infer hierarchical relations on four datasets (detailed statistics in Appendix G): MCG (Wang et al., 2015; Wu et al., 2012; Li et al., 2017), Hearst patterns (Hearst, 1992; Le et al., 2019), Word Net Noun (Christiane, 1998), and its Mammal sub-graph.
Researcher Affiliation Academia Tao Yu , Toni J.B. Liu , Albert Tseng, and Christopher De Sa Cornell University {ty367, jl3499, at676, cmd353}@cornell.edu
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Our code is available on github https://github.com/ydtydr/Shadow Cones
Open Datasets Yes This section showcases shadow cones ability to represent and infer hierarchical relations on four datasets (detailed statistics in Appendix G): MCG (Wang et al., 2015; Wu et al., 2012; Li et al., 2017), Hearst patterns (Hearst, 1992; Le et al., 2019), Word Net Noun (Christiane, 1998), and its Mammal sub-graph.
Dataset Splits Yes The remaining 10% of non-basic edges are evenly divided between the validation and test sets.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory specifications) used for running experiments.
Software Dependencies No The paper mentions using "HTorch (Yu et al., 2023) for optimization in various models of hyperbolic space. We use Riemannian SGD for Poincaré half-space model, and Riemannian Adam for Poincaré ball model". However, no specific version numbers are provided for HTorch or the optimizers, which are necessary for reproducibility.
Experiment Setup Yes For the margin parameters in shadow loss, we use γ2 = 0 consistently for all experiments. We tune γ1 and the learning rate in {0.01, 0.001, 0.0001}. For umbral cones, we tune the source radius r in {0.01, 0.05, 0.1, 0.2, 0.3}, empirically r = 0.05 during training gives the optimal performance when evaluated under a slightly larger radius r = 0.1. For penumbral-half-space cones, we tune the exponentiated height kh in {2, 5, 10, 20}... We set γ3 = 0.0001 consistently for all shadow cones. We use Riemannian SGD for Poincaré half-space model, and Riemannian Adam for Poincaré ball model. and A training batchsize of 16 is used for all datasets and models.