MaSDEs

Neural Stochastic Differential Games for Time-series Analysis

The PyTorch implementation for the paper titled "Neural Stochastic Differential Games for Time-series Analysis" by Sungwoo Park, Byoungwoo Park, Moontae Lee and Changhee Lee published at ICML 2023.

Park, S., Park, B., Lee, M., & Lee, C. (2023). Neural Stochastic Differential Games for Time-series Analysis. In International Conference on Machine Learning. PMLR.

<p align="center"> <img align="middle" src="https://github.com/LGAI-AML/MaSDEs/blob/main/imgs/masde_main.png" width="800" /> <br> <b> Figure 1. </b> Conceptual illustration: We utilizes game theory to model temporal dynamics of time-series data by extending conventional differential equation to the multi-agent counterpart for decomposing the time-series. </p> <p align="center"> <img align="middle" src="https://github.com/LGAI-AML/MaSDEs/blob/main/imgs/mackey.gif" width="400" /> <br> <b> Figure 2. </b> The highlighted region shows temporal aggregation level for each decision and the vertical axis in the past interval represents the average temporal aggregation of the agent's decision over the future interval. </p>

You can find more information and details about the project on [Paper] & [Project page].

Installation

This code is developed with Python3 and Pytorch. To set up an environment with the required packages, run

conda create -n masde python=3.8
conda activate masde
pip install -r requirements.txt

Training and Evaluation

Air Quality

python main.py -data_set air_quality -T_p 48 -T_o 36 -PI 12 -D 6 -ail

Physionet

python main.py -data_set physionet -T_p 48 -T_o 36 -PI 12 -D 36 -ail

Speech

python main.py -data_set speech -T_p 54 -T_o 43 -PI 11 -D 65 -ail