Robust and Conjugate Spatio-Temporal Gaussian Processes
Authors: William Laplante, Matias Altamirano, Andrew B. Duncan, Jeremias Knoblauch, Francois-Xavier Briol
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We study our method extensively in finance and weather forecasting applications, demonstrating that it provides a reliable approach to spatio-temporal modelling in the presence of outliers. In the remainder, we study the advantages of ST-RCGP on numerical examples. Throughout Section 4, we evaluate experiments based on root mean squared error RMSE(X, ˆy) := p EY p0( |X) [(Y ˆy)2] and evaluations of the negative log predictive distribution NLPD(X, ˆy, ˆσ) := EY p0( |X) log pϕ Y |ˆy, ˆσ2 on the test data. |
| Researcher Affiliation | Academia | 1Department of Physics and Astronomy, University College London, London, United Kingdom 2Department of Statistical Science, University College London, London, United Kingdom 3The Alan Turing Institute, London, United Kingdom 4Department of Mathematics, Imperial College London, London, United Kingdom. Correspondence to: William Laplante <EMAIL>. |
| Pseudocode | No | No explicit pseudocode or algorithm blocks are present in the paper. The methodology is described through mathematical equations and textual explanations in Section 3 and Proposition 3.1. |
| Open Source Code | Yes | The code to reproduce all experiments is available at https: //github.com/williamlaplante/ST-RCGP. |
| Open Datasets | Yes | The data is obtained from https://www.kaggle.com/c/ caltech-cs155-2020/data. The data we use is collected by the Climate Research Unit (see Harris et al., 2020) and measures temperature from 16/01/2022 to 16/12/2023 at ns = 571 locations, containing N = 11, 991 data points in total. The data is from the Climate Research Unit (CRU) and is available at https://crudata.uea.ac.uk/cru/data/ hrg/. The well-log dataset, first introduced by Ruanaidh & Fitzgerald (1996). |
| Dataset Splits | Yes | Hyperparameter optimisation is performed from 16/01/2022 to 30/09/2023, with later dates serving as test data. In October and November 2023, we perform in-sample predictions, and December 2023 is used for a one-month temperature forecast. RMSE and NLPD use 1000 test points around the induced crash that are not outliers. |
| Hardware Specification | Yes | All experiments are run on the CPU of a 2020 13-inch Mac Book Pro with M1 chip and 8GB of memory. |
| Software Dependencies | No | The paper mentions using 'Python's sklearn package' and the 'Bayes Newton package' (Wilkinson et al., 2023) but does not specify version numbers for these or any other software dependencies used in their implementation. |
| Experiment Setup | Yes | For both the STGP and ST-RCGP, we fit the data with the optimisation objective φ and use de-contaminated data (original data without outliers) for the objective. We use the Adam optimiser with 20 training steps and 0.4 learning rate. The two algorithms use a Mat ern 3/2 kernel. The ST-RCGP uses the adaptive centering and shrinking function from Section 3. |