---

Counterfactual Explanations for Multivariate Time Series

CONFETTI is a multi-objective method for generating counterfactual explanations for multivariate time series. It identifies the most influential subsequences, constructs a minimal perturbation, and optimizes it under multiple objectives to produce sparse, realistic, and confidence-increasing counterfactuals

CONFETTI is model-agnostic and works with any deep learning classifier, differentiable or not.

---

✨ Highlights

• Multi-objective optimization using NSGA-III
• Works for any Keras/Scikit-learn multivariate time series classifier
• Optional use of class activation maps for feature-weighted perturbations
• Generates multiple diverse counterfactuals per instance
• Parallelized counterfactual generation
• Built-in utilities for:
  • loading datasets
  • computing CAM weights
  • visualizing counterfactual explanations

---

🚀 Installation

PyPI Installation

pip install confetti-ts

Development Installation

git clone https://github.com/serval-uni-lu/confetti.git
cd confetti

uv venv
source .venv/bin/activate
uv pip install -e .

Requirements:

• Python 3.12+
• NumPy, pandas
• Keras 3.x
• TensorFlow
• Pymoo
• tslearn

All dependencies are handled automatically via pyproject.toml.

---

⚡ Quick Example

Below is a minimal end-to-end example based on the demo_confetti.ipynb notebook. It loads a trained model, prepares a dataset, and generates counterfactuals for a single instance.

from confetti import CONFETTI
from confetti.attribution import cam
from confetti.utils import load_multivariate_ts_from_csv
from confetti.visualizations import plot_counterfactual
import keras

# Load model
model_path = "examples/models/toy_fcn.keras"
model = keras.models.load_model(model_path)

# Load dataset in (n_samples, time_steps, n_features) format
X_train, y_train = load_multivariate_ts_from_csv("examples/data/toy_train.csv")
X_test, y_test   = load_multivariate_ts_from_csv("examples/data/toy_test.csv")

# Select instance to explain
instance = X_test[0:1]

# Generate CAM weights for training data (optional)
training_weights = cam(model, X_train)

# Initialize explainer
explainer = CONFETTI(model_path=model_path)

# Generate counterfactuals
results = explainer.generate_counterfactuals(
    instances_to_explain=instance,
    reference_data=X_train,
    reference_weights=training_weights,      # or None if not available
)

# Visualize the best counterfactual
plot_counterfactual(
    original=results[0].original_instance,
    counterfactual=results[0].best,
    cam_weights=results[0].feature_importance,
    cam_mode="heatmap",
    title="Counterfactual Explanation"
)

In the visualization:

• green curves represent the original instance
• red curves represent the counterfactual subsequence
• the heatmap corresponds to CAM scores of the nearest unlike neighbor

The alignment between CAM activation and the altered subsequence shows how CONFETTI uses attribution to target meaningful areas of the time series.

---

📚Documentation

The full documentation, including usage guides, API reference, and examples, is available at:

👉 https://confetti-ts.readthedocs.io/en/latest/

---

📄License

CONFETTI is released under the MIT License.

---

📝 Citing CONFETTI

If you use CONFETTI in your research, please consider citing the following paper:

@inproceedings{cetina2026counterfactual,
  title={Counterfactual Explainable AI (XAI) Method for Deep Learning-Based Multivariate Time Series Classification},
  author={Cetina, Alan Gabriel Paredes and Benguessoum, Kaouther and Lourenco, Raoni and Kubler, Sylvain},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={40},
  number={21},
  pages={17393--17400},
  year={2026}
}

To replicate the experiments described in the paper, use the paper branch of this repository. It contains the experiment scripts, model configurations, and dataset handling used in the publication.