Learning the Dynamics of Sparsely Observed Interacting Systems
Authors: Linus Bleistein, Adeline Fermanian, Anne-Sophie Jannot, Agathe Guilloux
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our theoretical results are illustrated by simulations which show that our method outperforms existing algorithms for recovering the full time series while being computationally cheap. We conclude by demonstrating its potential on real-world epidemiological data. We study the performance of Sig Lasso obtained by solving the optimization problem (8), where Ω(θ) is defined by Equation (11). All details are given in Appendix D. We compare Sig Lasso to a GRU and Neural CDE (Kidger et al., 2020). We measure the performance of the models with two metrics on a test set: the mean squared error for predicting the last observation point of the target paths and the L2 error for predicting the full path on a fine grid. |
| Researcher Affiliation | Collaboration | 1Inria Paris, F-75015 Paris, France 2Centre de Recherche des Cordeliers, INSERM, Universit e de Paris, Sorbonne Universit e, F-75006 Paris, France 3La MME, UEVE and UMR 8071, Paris Saclay University, F-91042, Evry, France 4MINES Paris Tech, PSL Research University, CBIO, F75006 Paris, France 5Institut Curie, PSL Research University, F-75005 Paris, France 6INSERM, U900, F-75005 Paris, France 7LOPF, Califrais Machine Learning Lab, Paris, France 8APHP, Paris, France. |
| Pseudocode | Yes | The Learn-And-Reconstruct algorithm is the generic algorithm used in our work... It is described in Algorithm 1. Algorithm 1 Learn-and-Reconstruct Algorithm. The algorithm infers for every individual in the test set a reconstructed time series ˆY i t. |
| Open Source Code | Yes | All proofs are postponed to the appendix and the code to reproduce the experiments is available at https://github.com/Linus Bleistein/Sig Lasso. |
| Open Datasets | Yes | We train our model to learn the dynamics linking population data related to mobility, vaccination, and weather, and the hospitalization growth rate (HGR) in each of the 9 metropolitan regions of France based on the data of Paireau et al. (2022). Hospital data was obtained from the SI-VIC database, the national inpatient surveillance system. Both SIDEP ( Syst eme d Information de D epistage Populationnel ) and VAC-SI datasets are publicly available. The mobility data was obtained from Google. The meteorological data was obtained from M et eo France. |
| Dataset Splits | No | The paper mentions that "Penalty strenght of the Sig Lasso is crossvalided using the internal implementation Lasso CV of scikit-learn", but it does not specify explicit dataset splits (e.g., percentages, k-fold count) for validation. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., CPU/GPU models, memory, cloud instance types) used for running the experiments. |
| Software Dependencies | No | The paper mentions software like "scikit-learn" (Pedregosa et al., 2011), "iisignature" (Reizenstein & Graham, 2020), and "torchcde" (Kidger et al., 2020) but does not provide specific version numbers for these dependencies, which is required for reproducible software setup. |
| Experiment Setup | Yes | The GRU is of width 128 and systematically trained with 100 epoches using a learning rate of 0.001. The NCDE is trained for 30 epochs with a learning rate of 0.001. It has 2 hidden layers of width 128, an intermediate Tanh( ) nonlinearity and a final linear readout. The depth of the signature is a hyperparameter chosen between 2 and 9 or 6 depending on the experiment. An intercept is added. |