Causal Discovery and Forecasting in Nonstationary Environments with State-Space Models
Authors: Biwei Huang, Kun Zhang, Mingming Gong, Clark Glymour
ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on synthetic and real-world data sets demonstrate the efficacy of the proposed methods. |
| Researcher Affiliation | Academia | 1Department of Philosophy, Carnegie Mellon University, Pittsburgh 2Department of Biomedical Informatics, University of Pittsburgh, Pittsburgh. |
| Pseudocode | Yes | Algorithm 1 Forecasting of YT +1 by Metropolis-Hastings |
| Open Source Code | No | The paper does not provide any statement or link for open-source code availability for the described methodology. |
| Open Datasets | Yes | We investigated the causal relationships between Gross Domestic Product (GDP), inflation, economic growth, and unemployment rate, with quarterly data from 1965 to 2017 in the USA 1. The data are normalized by subtracting the mean and dividing them by the standard deviation. 1Downloaded from https://www.theglobaleconomy.com/. |
| Dataset Splits | No | The paper mentions generating synthetic data and using real-world data for forecasting. For synthetic data, it notes 'simulated 10 values for the processes' for prediction. For real data, it forecasts 'from 2007 to 2017 with one-step prediction'. However, it does not explicitly provide detailed dataset splits (e.g., percentages, counts, or a specific split methodology) for training, validation, or testing that would allow reproduction of data partitioning. |
| Hardware Specification | No | The paper does not provide any specific hardware details used for running the experiments. |
| Software Dependencies | No | The paper mentions various algorithms and methods like Lasso, Kalman Filtering, Gaussian Process, SAEM, and CPF-AS, but it does not specify any software dependencies with version numbers (e.g., Python 3.x, PyTorch x.x, scikit-learn x.x.x). |
| Experiment Setup | Yes | In our methods, we randomly initialized the parameters and determined the causal graph by using a threshold (we simply used 0.05) for both the mean and variance of ˆbij,t; that is, if ˆbij = 1 T PT t=1 ˆbij,t < 0.05 and 1 T PT t=1(ˆbij,t ˆbij)2 < 0.05, we concluded that there is no edge from xj to xi. For CD-NOD, the kernel width was set empirically (Zhang et al., 2017), and the significance level was 0.05. Since both IB and MC methods need data from multiple domains, we segmented the data into non-overlapping domains with sample size 100 in each domain. ... Particularly, the window size for window-based Lasso was 100. |