Nonlinear Feature Diffusion on Hypergraphs
Authors: Konstantin Prokopchik, Austin R Benson, Francesco Tudisco
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We now evaluate our method on several real-world hypergraph datasets (Table 1). We compare our method to six baselines. |
| Researcher Affiliation | Academia | 1Gran Sasso Science Institute, L Aquila, Italy 2Cornell University, New York, USA. Correspondence to: Konstantin Prokopchik <konstantin.prokopchik@gssi.it>, Austin R. Benson <arb@cs.cornell.edu >, Francesco Tudisco <francesco.tudisco@gssi.it>. |
| Pseudocode | Yes | The overall SSL algorithm is detailed in Algorithm 1. ... Algorithm 1 Hyper ND: Nonlinear Hypergraph Diffusion |
| Open Source Code | Yes | The code implementing the experiments is available at https://github.com/compile-gssi-lab/Hyper ND. |
| Open Datasets | Yes | We use five co-citation and co-authorship hypergraphs: Cora co-authorship, Cora cocitation, Citeseer, Pubmed (Sen et al., 2008) and DBLP (Rossi & Ahmed, 2015). We also consider a foodweb hypergraph... (foo). |
| Dataset Splits | Yes | For our method and HTV, which have no training phase, we run 5-fold cross validation with label-balanced 50/50 splits to choose α from {0.1, 0.2, . . . , 0.9} and p from {1, 2, 3, 5, 10}. Precisely, we split the data into labeled and unlabeled points. We split the labeled points into training and validation sets of equal size (label-balanced 50/50 splits) and we choose the parameters based on the average performance on the validation set over 5 random repeats. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory specifications) used for running its experiments. It only mentions 'Execution time' comparisons. |
| Software Dependencies | No | The paper mentions "BLAS routines from, e.g., Num Py" but does not specify version numbers for any software dependencies, which is required for reproducibility. |
| Experiment Setup | Yes | For all of the neural network-based models, we use two layers and 200 training epochs, following (Yadati et al., 2019) and (Feng et al., 2019). ... For our method and HTV, which have no training phase, we run 5-fold cross validation with label-balanced 50/50 splits to choose α from {0.1, 0.2, . . . , 0.9} and p from {1, 2, 3, 5, 10}. ... All the datasets we use here have nonnegative input embedding [Y X] which we preprocess via label smoothing as in (12), with ε = 1e 6. |