xai4tsc 1.1.0

An evaluation framework for eXplainable AI (XAI) methods applied to Time Series Classification (TSC).

XAI4TSC

An evaluation framework for eXplainable AI (XAI) methods applied to Time Series Classification (TSC), developed by the TimeXAI Research Group.

XAI4TSC has two independent use cases:

• Standalone experiment runner: Clone the repo, choose or adapt a YAML config, run experiments from the command line.
• Importable Python package: pip install xai4tsc and use the public API in your own code, notebooks, or scripts.

The documentation can be found here: https://timexaigroup.github.io/XAI4TSC/

---

Installation

Standalone (experiment runner)

We use Poetry for package management inside a Conda environment:

1. Clone the repository:

   git clone https://github.com/TimeXAI-group/XAI4TSC.git
   cd XAI4TSC
   

2. Create a python environment (choose one of the following):
   2.1. Using python:

   python -m venv .venv			 # Use python 3.12 or 3.13
   source .venv/bin/activate
   

   2.2. Using conda:

   conda env create          # picks up environment.yml
   conda activate xai4tsc
   

3. Install the dependencies (choose one of the following):
   3.1. Using poetry:

   pip install poetry # only needed if the local python .venv is used
   poetry install
   

   3.2. Using pip and PyPI:

   pip install xai4tsc
   

   3.3. Using pip and a local installation:

   pip install -e . 
   

Package only

pip install xai4tsc

Or, for a local/editable install from a clone:

pip install -e PATH/TO/REPOSITORY

---

Project layout

xai4tsc/
├── experiment_runner/          # Standalone CLI — owns all config and orchestration logic
│   ├── main.py                 # Entry point: python -m experiment_runner.main --conf ...
│   ├── config.py               # Config loading, validation, defaults
│   ├── cache.py                # Runner-level split caching helpers
│   ├── explain.py              # Runner adapter for xai4tsc.xai
│   ├── evaluate.py             # Runner adapter for xai4tsc.evaluation
│   ├── log_setup.py            # Logging setup for the standalone runner
│   └── configs/
│       ├── master.yaml         # Annotated reference — all available options and defaults
│       ├── example.yaml        # Go-to demo: synthetic data, time-domain explainers + metrics
│       ├── example_frequency.yaml  # Frequency / time-frequency showcase (FreqRISE, freq metrics)
│       ├── ucr_benchmark.yaml  # Full UCR sweep (LeNet + Integrated Gradients + Complexity)
│       └── uea_benchmark.yaml  # Full UEA sweep (skips OOM-risky datasets)
│
├── src/xai4tsc/                # Importable package
│   ├── data/
│   │   ├── datasets.py         # UcrUeaDataset, LocalDataset, SyntheticDataset
│   │   └── ...                 # base classes, data loaders
│   ├── models/                 # ModelBase, built-in models, registry
│   ├── xai/                    # Explainer ABC, generate_explanation(), built-in explainers
│   ├── evaluation/             # evaluate(), Quantus metric registry
│   └── utils/                  # Shared utilities (dict_to_args, merge_dicts, rescale_array, plot)
│
└── tests/
    ├── conftest.py             # Session-scoped fixtures (GunPoint download, split, model)
    ├── fixtures/
    │   └── test_config.yaml    # Minimal runner config for integration tests
    ├── unit/                   # Fast, no-I/O tests  (@pytest.mark.unit)
    └── integration/            # Full pipeline tests (@pytest.mark.integration)

---

Running an experiment

Run the runner as a module from the repository root:

python -m experiment_runner.main --conf experiment_runner/configs/example.yaml
python -m experiment_runner.main --conf experiment_runner/configs/example.yaml --debug

Four ready-to-run configs ship under experiment_runner/configs/:

• example.yaml — the go-to demo (and the default when --conf is omitted): the synthetic freq_shapes dataset, two models, and time-domain explainers and metrics. Needs no download.
• example_frequency.yaml — the same dataset explained with FreqRISE in the frequency and time-frequency domains, scored with the frequency metrics.
• ucr_benchmark.yaml / uea_benchmark.yaml — full-archive sweeps with a minimal model/explainer/metric stack.

Edit or copy any of them to configure datasets, models, explainers, and metrics. master.yaml contains annotated documentation for every available option.

---

Testing

The test suite uses two pytest markers to separate fast unit tests from slow integration tests that require a live dataset and a training run.

# Unit tests only — fast, no I/O, no training
pytest -m unit

# Integration tests only — downloads GunPoint on first run, trains a model
pytest -m integration

# Full suite
pytest

GunPoint is downloaded automatically on the first integration run and cached in tests/cache/ (override with the XAI4TSC_TEST_CACHE environment variable).

---

Using xai4tsc as a package

import xai4tsc
from xai4tsc.data import load_dataset, LocalDataset
from xai4tsc.models.models import load_model
from xai4tsc.xai.explain import generate_explanation
from xai4tsc.evaluation.evaluate import evaluate

# ── Load data (UCR download or local numpy files) ─────────────────────────────
ds = load_dataset("GunPoint")                           # UcrUeaDataset — downloads on first use
# ds = LocalDataset("/path/to/data", name="MyDataset")  # local data.npy + labels.json

splits, encoder = ds.split(
    train_split=0.8, val_split=0.1, random_state=42, encode="label"
)
train_data, train_labels, _ = splits[0]
test_data,  test_labels,  _ = splits[1]

# ── Train a model ─────────────────────────────────────────────────────────────
model = load_model(
    {"model": "FCN", "init_params": {"in_channels": 1, "num_classes": 2}},
    device="cpu",
)
model.train_model(
    train_data, train_labels,
    hyperparams={"epochs": 50, "batchsize": 32, "loss_func": "CrossEntropy",
                 "optimizer": "adam", "learn_rate": 0.001, "patience": 10},
    save_path="results",   # best checkpoint + training plots land here
)

# ── Generate explanations ─────────────────────────────────────────────────────
exp = generate_explanation(
    method="integrated_gradients",
    model=model,
    data=test_data,
    labels=test_labels,
    encoder=encoder,
    indices=[0, 1, 2],
    device="cpu",
)
# exp.exp_values — numpy array, same shape as test_data[[0, 1, 2]]

# ── Evaluate ──────────────────────────────────────────────────────────────────
score = evaluate(
    model=model,
    metric="Complexity",
    explanation=exp,
    data=test_data[exp.indices],
    labels=test_labels[exp.indices],
    metric_class_params={"normalise": True, "abs": True, "disable_warnings": True},
    device="cpu",
)

# ── Register a custom explainer ───────────────────────────────────────────────
from captum.attr import Saliency
from xai4tsc import GradientExplainer, register_explainer

class SaliencyExplainer(GradientExplainer):
    def _get_captum_attribution(self, model):
        return Saliency(model)

register_explainer("saliency", SaliencyExplainer)
exp2 = generate_explanation(
    "saliency", model=model, data=test_data,
    labels=test_labels, encoder=encoder, indices=[0], device="cpu",
)

---

Built on

• Computational backends: PyTorch, scikit-learn
• Explanation backend: Captum
• Evaluation backend: Quantus

---

Contributing

Contributions are welcome. See CONTRIBUTING.md for the development setup, code style and quality gate, testing conventions, and the pattern for adding a new model, explainer, or metric. By participating you agree to the Code of Conduct. The project is released under the MIT License.

---

Disclaimer

XAI4TSC bundles clean-room re-implementations of models, explainers, and metrics from the research literature, alongside thin wrappers around third-party libraries (Captum, Quantus). These implementations have not been verified by the original authors and may differ from the source papers or reference code in ways that affect results.

Evaluation scores are an empirical, sometimes contested, proxy for explanation quality — treat them as guidance, not ground truth. When using XAI4TSC for research, cite the original papers, state which implementation you used, and, where possible, validate against a reference implementation. See the full disclaimer in the documentation for details.