Towards a Rigorous Evaluation of XAI Methods on Time Series

This project explores Explainable AI (XAI) techniques for 1D time series classification using deep learning models. It applies three popular attribution methods from the Captum library to analyze and compare the decision-making processes of two models: SimpleCNN and ResNet1D.

We use Deletion and Insertion Curves to assess how accurately each attribution method reflects the model’s reasoning.

---

📂 Dataset

The dataset used is a 1D time series classification dataset.

Example:
Dataset: FordA (from UCR Time Series Classification Archive)

• 3601 training samples, 1320 testing samples
• Input shape: (1, 500)
• Number of classes: 2
• Preprocessing: Normalized to range [0, 1]

---

Models

Two different 1D models are implemented and trained on the dataset:

1. SimpleCNN

• A lightweight convolutional neural network.
• Architecture:
  • Conv1D → ReLU → MaxPool1D → Dropout → Linear → Softmax

2. ResNet1D

• A 1D version of the ResNet architecture.
• Uses residual blocks to capture deeper temporal features.

These models are assumed to be trained and available as model_cnn and model_resnet.

---

📌 Explainability Methods

Three attribution methods are used from the Captum library:

Method	Description
IntegratedGradients	Computes average gradients along a straight-line path from a baseline to the input.
Saliency	Computes the gradient of the output with respect to the input features.
DeepLift	Compares input activation to a reference and propagates differences.

Each method helps determine which parts of the input most influenced the model's prediction.

---

📉 Deletion and Insertion Curves

To evaluate the faithfulness of attributions, the project uses Deletion and Insertion metrics:

• Deletion: Gradually remove the most important features (by zeroing them) and observe the drop in prediction confidence.
• Insertion: Start from a blank input and gradually add the most important features to observe the confidence rise.

Each attribution method is tested on both models using a single test sample.

Visualization

• A 2×3 grid of plots is generated:
  • 2 rows: Models (SimpleCNN, ResNet1D)
  • 3 columns: Attribution methods

Each subplot shows:

• X-axis: Fraction of features perturbed
• Y-axis: Prediction probability for the correct class
• Red line: Deletion curve
• Green line: Insertion curve

---

📁 File Structure

File/Function	Description
models_dict	Dictionary with model names and instances
attr_methods	Dictionary of Captum attribution method classes
deletion_insertion_curve(...)	Computes the Deletion or Insertion scores
plot_deletion_insertion_ax(...)	Plots a single subplot for one model-method pair
plt.subplots(...)	Generates all 6 plots for visual comparison

---

▶️ How to Run

1. Install dependencies:

pip install torch torchvision captum matplotlib numpy

2. Load the dataset (ECG200) and define your models.

3. Run the attribution and evaluation script.

---

📎 File Overview

• model.py: Defines SimpleCNN and ResNet1D architectures
• attributions.py: Applies Captum-based attribution methods
• deletion_insertion.py: Contains deletion & insertion evaluation logic
• main.py: Loads data, computes attributions, and plots results

---

📌 References

• UCR Time Series Archive: http://www.timeseriesclassification.com/
• Captum Library: https://captum.ai/
• Paper: Towards a Rigorous Evaluation of XAI Methods on Time Series
• Paper: Axiomatic Attribution for Deep Networks (Integrated Gradients)
• Paper: Bayesian XAI methods towards a robustness-centric approach to Deep Learning: an ABIDE I study

---

🧪 Future Work

• Add support for SHAP & LIME
• Use multi-class datasets from UCR
• Implement faithfulness metrics like AUC or area between insertion & deletion curves