Frequency-aware Generative Models for Multivariate Time Series Imputation
Authors: XINYU YANG, Yu Sun, Yuan xiaojie, Xinyang Chen
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | This section experimentally evaluates both the imputation effectiveness and the improvement of real downstream applications for our FGTI, against various competing methods. All experiments are performed on a machine with Intel Core 3.0GHz i9 CPU, NVIDIA Ge Force RTX 3090 24GB GPU, and 64GB RAM. The source code and datasets are available online [1]. |
| Researcher Affiliation | Academia | Xinyu Yang1, Yu Sun1 , Xiaojie Yuan1, Xinyang Chen2 1College of Computer Science, DISSec, Nankai University, China {yangxinyu@dbis.,sunyu@,yuanxj@}nankai.edu.cn 2School of Computer Science and Technology, Harbin Institute of Technology, Shenzhen, China chenxinyang@hit.edu.cn |
| Pseudocode | Yes | Algorithm 1 Training process of FGTI implemented by the diffusion model Input: Incomplete time series X, the number of diffusion step T Output: Optimized denoising network ϵθ( ) |
| Open Source Code | Yes | The source code and datasets are available online [1]. [1] https://github.com/FGTI2024/FGTI24. |
| Open Datasets | Yes | We employ three real time series datasets with real-world missing values. KDD [6] collects 8,034 meteorological and air quality readings of nine stations from January 30, 2017 to January 31, 2018 in Beijing, with 4.46% real missing values. Guangzhou [10] records traffic speeds per ten minutes on 214 anonymous roads in Guangzhou from August 1, 2016 to September 30, 2016. There are 1.29% real missing values in the dataset. Physio Net [42] contains 37 measurement readings from 11,988 patients within 48 hours of the ICU admission. 79.71% measurements are missing in the dataset, and 1,707 patients died after 48 hours of the ICU admission. |
| Dataset Splits | No | The paper mentions 'imputation target' and 'missing values' for evaluation, and discusses 'mask ratios' for training the denoising network, but it does not specify explicit train/validation/test dataset splits with percentages or sample counts. |
| Hardware Specification | Yes | All experiments are performed on a machine with Intel Core 3.0GHz i9 CPU, NVIDIA Ge Force RTX 3090 24GB GPU, and 64GB RAM. |
| Software Dependencies | No | The paper mentions specific methods and models like 'SAITS', 'MIWAE', 'GPVAE', 'CSDI', 'PRi STI', but it does not list the specific version numbers of any software libraries, frameworks, or programming languages used for implementation. |
| Experiment Setup | Yes | For methods in which the authors recommend parameters such as SAITS, MIWAE, GPVAE, CSDI and PRi STI, we use these parameters as suggested. The other methods are also configured in a best-effort fashion by iteratively choosing good parameters. We set the cutoff frequency F of the high-frequency filter to 0.3 and set the number of maximum magnitude frequency κ of the dominant-frequency filter to 10. In addition, for other settings related to the diffusion model, we adopt hyperparameters recommended by the existing well-established models [44; 24]. |