Stationarity without mean reversion in improper Gaussian processes
Authors: Luca Ambrogioni
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | By analyzing both synthetic and real data, we demonstrate that these non-positive kernels solve some known pathologies of mean reverting GP regression while retaining most of the favorable properties of ordinary smooth stationary kernels. |
| Researcher Affiliation | Academia | 1Donders Institute for Brain, Cognition and Behaviour, Radboud University. |
| Pseudocode | No | The paper does not contain any pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | No | The paper does not provide any explicit statement about making its source code available or include a link to a code repository. |
| Open Datasets | Yes | Specifically, we test on regression problems from the UCI database (Leisch and Dimitriadou, 2021; D.J. et al., 1998). ... We used the following datasets: 3droad (n = 434874, d = 3); autompg (n=392, d=7); bike (n=17379, d=17); concreteslump (n=103, d=7); energy (n=768; d=8); forest (n=517, d=12); houseelectric(n=2049280, d=11); keggdirected (n=48827, d=20); kin40k(n=40000, d=8); parkinsons (n=5875, d=20); pol (n = 15000, d=26); pumadyn32nm (n=8192; d=32); slice (n=53500, d=385); solar (n=1066, d=10); stock (n=536; d=11); yacht (n=308, d=6); airfoil (n=1503, d=5); autos (n=159, d=25); breastcancer (n=194, d=33); buzz (n=583250, d=77); concrete (n=1030, d=8); elevators (n=16599, d=18); fertility (n=100, d=9); gas (n=2565, d=128); housing (n=506, d=13); keggundirected (n=63608, d=27); machine (n=209,d=7); pendulum (n=630, d=9); protein (n=45730, d=9); servo (n=167, d=4); skillcraft (n=3338, d=19); sml (n=4137, d=26); song (n=515345, d=90); tamielectric (n=45781, d=3); wine (n=1599, d=11). |
| Dataset Splits | No | The paper does not provide explicit details about a separate validation set or a standard train/validation/test split. It mentions training, testing, and bootstrapping/resampling, but not distinct validation splits. |
| Hardware Specification | No | The paper does not provide any specific details about the hardware used for running the experiments (e.g., GPU models, CPU types, memory). |
| Software Dependencies | No | The paper mentions using the "yfinance (yahoo finance) python package" for stock prices, but it does not specify version numbers for this or any other software dependencies, which is required for reproducibility. |
| Experiment Setup | Yes | The length scale was 0.2 for both models. ... The length scale was selected by optimizing the marginal likelihood conditioned on one data point (see rightmost panel). ... noise level σ = 0.05. ... Since the actual marginal likelihood is divergent in improper models, the hyperparameters were optimized by minimizing the marginal likelihood conditional to one randomly selected observation. ... All kernels were parameterized by a length scale and a noise std parameter, which were optimized by maximizing the log-marginal likelihood conditional to one, randomly chosen, datapoint. |