Representation Tradeoffs for Hyperbolic Embeddings
Authors: Frederic Sala, Chris De Sa, Albert Gu, Christopher Re
ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the proposed approaches and compare against existing methods. We hypothesize that for tree-like data, the combinatorial construction offers the best performance. For general data, we expect h-MDS to produce the lowest distortion, while it may have low MAP due to precision limitations. We anticipate that dimension is a critical factor (outside of the combinatorial construction). In the appendix, we report on additional datasets, combinatorial construction parameters, and the effect of hyperparameters. |
| Researcher Affiliation | Academia | 1Department of Computer Science, Stanford University 2Department of Computer Science, Cornell University. |
| Pseudocode | Yes | Algorithm 1 Sarkar s Construction |
| Open Source Code | No | The paper mentions a "Py Torch-based implementation" but does not provide any link or explicit statement about the availability of the source code for their methodology. |
| Open Datasets | Yes | Trees include fully-balanced and phylogenetic trees expressing genetic heritage (Hofbauer et al., 2016), available at Sanderson et al. (1994). Nearly tree-like hierarchies include the Word Net hypernym graph (the largest connected component from Nickel & Kiela (2017)) and a graph of Ph.D. advisor-advisee relationships (De Nooy et al., 2011). Also included are datasets that vary in their tree nearness, such as disease relationships (Goh et al., 2007) and protein interactions (Jeong et al., 2001), both available from Rossi & Ahmed (2015). |
| Dataset Splits | No | The paper does not provide specific training/validation/test dataset splits (e.g., percentages or sample counts) for reproducibility. It mentions sampling the distance matrix for incomplete information but not standard data splits. |
| Hardware Specification | No | The paper does not specify any hardware details (e.g., GPU models, CPU types, memory) used for running the experiments. It only mentions a "Py Torch-based implementation". |
| Software Dependencies | No | The paper mentions using "Py Torch" for implementation but does not specify any version numbers for PyTorch or other software dependencies. |
| Experiment Setup | Yes | Combinatorial embeddings into H2 use the ε = 0.1 precision setting; others are considered in the Appendix. We performed h-MDS in floating point precision. [...] To evaluate our algorithm s ability to deal with incomplete information, we sample the distance matrix at a ratio of nonedges to edges at 10 : 1 following Nickel & Kiela (2017). |