Evaluating Robustness and Uncertainty of Graph Models Under Structural Distributional Shifts

Authors: Gleb Bazhenov, Denis Kuznedelev, Andrey Malinin, Artem Babenko, Liudmila Prokhorenkova

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In our experiments, we thoroughly evaluate the proposed distributional shifts and show that they can be quite challenging for existing graph models.
Researcher Affiliation Collaboration Gleb Bazhenov HSE University, Yandex Research Denis Kuznedelev Yandex Research, Skoltech Andrey Malinin Isomorphic Labs Artem Babenko Yandex Research, HSE University Liudmila Prokhorenkova Yandex Research
Pseudocode No The paper describes algorithms and methods but does not include any pseudocode or clearly labeled algorithm blocks.
Open Source Code Yes whereas the remaining methods are implemented in our custom experimental framework and can be found in our repository.5 (Footnote 5: Link to our Git Hub repository)
Open Datasets Yes For our experiments, we pick the following seven homophilous datasets that are commonly used in the literature: three citation networks, including Cora ML, Cite Seer [26, 9, 8, 31], and Pub Med [28], two co-authorship graphs Coauthor Physics and Coauthor CS [32], and two co-purchase datasets Amazon Photo and Amazon Computer [25, 32]. Moreover, we consider OGB-Products, a large-scale dataset from the OGB benchmark.
Dataset Splits Yes The half of nodes with the smallest values of σi are considered to be ID and split into Train, Valid-In, and Test-In uniformly at random in proportion 30% : 10% : 10%. The second half contains the remaining OOD nodes and is split into Valid-Out and Test-Out in the ascending order of σi in proportion 10% : 40%.
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models or types of computing resources used for experiments.
Software Dependencies No The paper mentions software components like GCN [16], SAGE [11], and Adam optimizer [14], but it does not specify their version numbers for reproducibility.
Experiment Setup Yes The hidden dimension of our baseline architecture is 256, and the dropout between hidden layers is p = 0.2. We exploit a standard Adam optimizer [14] with a learning rate of 0.0003 and a weight decay of 0.00001.