Intervention-Aware Time Series Modeling: Capturing and Evaluating Feature Dependencies

AAAI 2026 — Accepted

This repository contains the official implementation for the AAAI 2026 paper "Intervention-Aware Time Series Modeling: Capturing and Evaluating Feature Dependencies". The codebase provides tools for intervention-aware conditional VAE modeling, feature adjustment experiments, and evaluation of feature dependencies in multivariate time series.

Highlights

• Intervention-aware generative modeling for time series (CFORVAE)
• Feature adjustment/intervention experiments to analyze dependencies
• Forecasting and reconstruction tasks with reproducible scripts
• Benchmark datasets

Repository Structure

• CFORVAE.py, train_VAE.py, ts_reconstruct.py — VAE model - training & reconstruction
• train_forecast.py, test_forecast.py — forecasting models train and test
• st_sample_prediction.py - streamlit script for UI based comparison
• pages/ - streamlit pages
• models/, layers/, utils/ — model components and utilities
• dataset/ — datasets (AirQualityUCI, ETTh1, etc.)
• weights_forecast/, weights_recon/ — model weights folder

Quick Start

3. Train reconstruction models: python train_VAE.py --data pm25 --model CFORVAE
4. Train forecasting models: python train_forecast.py
5. Run evaluation/intervention UI: python -m streamlit run st_sample_prediction.py

Sample Feature Adjustment

Figure: Feature-Level Response to Targeted Interventions on the PM2.5 Dataset. The second column shows a feature that is correlated with the adjusted one (first column), where changes are expected to propagate. The third column displays an uncorrelated feature, where little to no response is expected

Citation

Will be available after it is officially published!