Graph Neural Network Bandits
Authors: Parnian Kassraie, Andreas Krause, Ilija Bogunovic
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5 Experiments We create synthetic datasets which may be of independent interest and can be used for evaluating and benchmarking machine learning algorithms on graph domains. Each dataset is constructed from a finite graph domain together with a reward function. ... Regret Experiments. We assess the performance of the algorithms on bandit optimization tasks over different domains. In Figure 1, we show the inference cumulative regret ... Figure 1 presents the results: GNN-PE consistently outperforms the other methods. |
| Researcher Affiliation | Academia | Parnian Kassraie ETH Zurich pkassraie@ethz.ch Andreas Krause ETH Zurich krausea@ethz.ch Ilija Bogunovic University College London i.bogunovic@ucl.ac.uk |
| Pseudocode | Yes | We introduce GNN-Phased Elimination (GNN-PE; see Algorithm 1) that consists of episodes of pure exploration over a set of plausible maximizer graphs, similar to [7, 30]. |
| Open Source Code | Yes | 3. If you ran experiments... (a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] Appendix D.2 includes the practical details and the instructions required to produce the results. |
| Open Datasets | Yes | We create synthetic datasets which may be of independent interest and can be used for evaluating and benchmarking machine learning algorithms on graph domains. ... instruction to create similar random datasets is also given. |
| Dataset Splits | No | The paper describes generating synthetic datasets for bandit problems, where data is observed sequentially. It does not specify fixed train/validation/test splits with percentages or counts in the traditional supervised learning sense. |
| Hardware Specification | No | The paper states: 'About 5 hour per CPU core on 100 internal cluster nodes.' This indicates the type of resource (CPU core, internal cluster nodes) but lacks specific model numbers for CPUs, GPUs, or detailed memory specifications. |
| Software Dependencies | No | The paper states: 'We use the Neural Tangents and Py Torch libraries. Both cited.' However, no specific version numbers for these libraries or other software dependencies are provided. |
| Experiment Setup | Yes | Experiment Setup. ... We always set width m = 2048 and layers L = 2, for every type of network architecture. ... To configure these algorithms, we only tune λ and β = βt... The number of gradient descent steps J is chosen adaptively, so that it depends on the number of observations, i.e., J = 10 2 log(t + 1). We use Adam optimizer with default parameters, and set the learning rate η to 5 10 3. The batch size is set to 256. |