Transformers Generalize DeepSets and Can be Extended to Graphs & Hypergraphs

Authors: Jinwoo Kim, Saeyoon Oh, Seunghoon Hong

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we demonstrate the capability of higher-order Transformers on a variety of tasks including synthetic data, large-scale graph regression, and set-to-(hyper)graph prediction. Specifically, we use a synthetic node classification dataset from Gu et. al. (2020) [11], a molecular graph regression dataset from Hu et. al. (2021) [17], two set-to-graph prediction datasets from Serviansky et. al. (2020) [29], and three hyperedge prediction datasets used in Zhang et. al. (2020) [36].
Researcher Affiliation Academia Jinwoo Kim, Saeyoon Oh, Seunghoon Hong School of Computing, KAIST {jinwoo-kim, saeyoon17, seunghoon.hong}@kaist.ac.kr
Pseudocode No The paper does not contain any pseudocode or clearly labeled algorithm blocks.
Open Source Code Yes Our implementation is available at https://github.com/jw9730/hot.
Open Datasets Yes Specifically, we use a synthetic node classification dataset from Gu et. al. (2020) [11], a molecular graph regression dataset from Hu et. al. (2021) [17], two set-to-graph prediction datasets from Serviansky et. al. (2020) [29], and three hyperedge prediction datasets used in Zhang et. al. (2020) [36].
Dataset Splits Yes We used training and test sets with chains of length 20 and 200 respectively. ... As test data is unavailable, we report the Mean Absolute Error (MAE) on validation dataset.
Hardware Specification Yes Plots are shown until each model runs into out-of-memory error on a RTX 6000 GPU with 22GB.
Software Dependencies No The paper does not explicitly provide a list of software dependencies with specific version numbers (e.g., 'Python 3.8, PyTorch 1.9, and CUDA 11.1').
Experiment Setup Yes Details including the datasets and hyperparameters can be found in Appendix A.2.