Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
How Expressive are Transformers in Spectral Domain for Graphs?
Authors: Anson Bastos, Abhishek Nadgeri, Kuldeep Singh, Hiroki Kanezashi, Toyotaro Suzumura, Isaiah Onando Mulang'
TMLR 2022 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical results suggest that Fe TA provides homogeneous performance gain against vanilla transformer across all tasks on standard benchmarks and can easily be extended to GNN-based models with low-pass characteristics (e.g., GAT). ... Extensive experiments with position encodings (PE) show that none of the considered state-of-the-art PEs proposed for transformer work are agnostic to the graph dataset. ... We study the efficacy of Fe TA with extensive experiments on standard benchmark datasets of graph classification/regression and node classification resulting in superior performance compared to vanilla transformers. |
| Researcher Affiliation | Collaboration | Anson Bastos EMAIL Indian Institute of Technology Hyderabad, India; Abhishek Nadgeri EMAIL RWTH Aachen and Zerotha Research, Germany; Kuldeep Singh Zerotha Research and Cerence Gmb H Aachen, Germany; Hiroki Kanezashi The University of Tokyo Tokyo, Japan; Toyotaro Suzumura The University of Tokyo Tokyo, Japan; Isaiah Onando Mulang IBM Research Kenya, Africa |
| Pseudocode | No | The paper describes methods and equations in prose and mathematical notation but does not contain any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain an explicit statement about releasing source code or a link to a code repository for the methodology described. |
| Open Datasets | Yes | We use widely popular datasets (Mialon et al., 2021; Hu et al., 2020): MUTAG, NCI1, and the OGB-Mol HIV for graph classification; PATTERN and CLUSTER for node classification; and ZINC for graph regression task (details in A.4). ... MUTAG (Morris et al., 2020), NCI1 (Morris et al., 2020) and, the ogbg-Mol HIV (Hu et al., 2020) dataset, for node classification we use the PATTERN and CLUSTER datasets (Dwivedi et al., 2020) and for graph regression we run our method on the ZINC (Dwivedi et al., 2020) dataset. |
| Dataset Splits | Yes | We borrow experiment settings and baseline values from (Dwivedi & Bresson, 2020; Mialon et al., 2021; Kreuzer et al., 2021). |
| Hardware Specification | Yes | Table 7: Computational details used for the datasets on the Fe TA-Base setting. Time is in seconds per epoch. MUTAG Geforce P8 8; NCI1 Geforce P8 8; Molhiv Tesla V100 16; PATTERN Tesla V100 16; CLUSTER Tesla V100 16; ZINC Tesla V100 16 |
| Software Dependencies | No | The paper discusses various models and techniques but does not explicitly list specific software dependencies with version numbers. |
| Experiment Setup | Yes | Table 8: Model architecture parameters of Fe TA-Base and Vanilla-Transformer. ... Table 9: Model architecture parameters of Fe TA with position embedding from Graphi T and original Graphi T model. ... Table 10: Parameters of Fe TA with position embedding from SAN compared with original SAN. ... For the configurations Fe TA-Base, Fe TA+Lap E, Fe TA+3RW and Fe TA+GCKN+3RW we used the default hyper-parameters provided by Graphi T (Mialon et al., 2021). |