Unsupervised Parameter-free Simplicial Representation Learning with Scattering Transforms

Authors: Hiren Madhu, Sravanthi Gurugubelli, Sundeep Prabhakar Chepuri

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

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical evaluations demonstrate that SSNs outperform existing simplicial or graph neural models in many tasks like node classification, simplicial closure, graph classification, trajectory prediction, and simplex prediction while being computationally efficient. In this section, we empirically evaluate SSN, assessing its performance across five downstream tasks: node classification, graph classification, simplicial closure prediction, trajectory prediction, and simplex prediction.
Researcher Affiliation Academia 1 Indian Institute of Science, India. Correspondence to: Hiren Madhu <hirenmadhu@iisc.ac.in>, Sravanthi Gurugubelli <sravanthig@iisc.ac.in>.
Pseudocode No The paper describes its method using prose and mathematical equations, but it does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes The code to run the experiments is available at https://github.com/ Hiren Madhu/SSN.
Open Datasets Yes The datasets used in the experiments and the number of simplices of various orders in them are summarized in Table 6. These include high-school, primary-school, senate-bills, email-Enron, madison-reviews, algebra-questions, and geometry-questions. ... We carry out node classification on publicly available datasets, namely the primary-school, high-school, and senate-bills datasets (Philip et al., 2021). ... Our experiments are conducted on the PROTEINS, NCI1, REDDIT-B, REDDIT-M, and IMDB-B datasets from the TUDatasets repository (Morris et al., 2020).
Dataset Splits Yes We conduct an experiment where we first extract features using SSN and then partition the data into different splits, such as 20 80, 40 60, 60 40, and 80 20. Here, the split notation a-b represents a% of the data used for training and b% for testing the linear classifier trained at the end. ... The data is temporally split, with the first 80% used for training the encoder and the remaining 20% for inference.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., CPU, GPU models, or cloud instances) used for running the experiments.
Software Dependencies No The paper mentions the use of 'Adam optimizer' but does not provide specific version numbers for any software dependencies or libraries.
Experiment Setup Yes To ensure a fair comparison, we standardize the use of two message passing layers for all neural models, which include SNNs and GNNs. However, we adjust the hidden dimensions and learning rate for each encoder to optimize performance. The encoder parameters are optimized using the Adam optimizer. For SSN, we employ two scattering layers and experiment with different Js, ranging from 1 to 10. We present the best performance achieved using all Js.