How aitrics/delta-xai works
Delta-XAI is a time-series explainability research framework. Its documented workflow loads a MIMIC-III dataset, initializes and loads a prediction model, selects an explainer, and computes feature attributions. The graph also preserves model training, generator-based explainers, TimeX variants, and visualization. Edges follow the README or shown source references; unsampled implementation details are not treated as verified calls.
Main entry points: run.py Explanation runner · modeltrainer.py Model trainer · base.py Explainer base · fit.py FIT explainer
Components 22
Experiment workflow
- Explanation runner
run.py - Experiment configuration
utils/config.py
Data handling
- Dataset loaders
data/dataloader.py - Data preprocessing
data/data_preprocess.py - ICU mortality data
data/icu_mortality.py
Predictive models
- Model trainer
modeltrainer.py - Model wrappers
models/base.py - Sequence models
models/lstm.py - CNN models
models/cnn.py - Transformer models
models/transformer.py - Interpretability layers
models/layers.py - Feature generators
models/generator.py
Attribution methods
- Explainer base
explainer/base.py - FIT explainer
explainer/fit.py - WinIT explainer
explainer/winit.py - SWING explainer
explainer/swing.py - TimeX explainer
explainer/timex.py - TimeX++ explainer
explainer/timexpp.py - Other explainers
explainer
Results and visualization
- Attribution visualization
utils/plot.py - Masking utilities
utils/masker.py
Other components
- Researcher
Connections 22
- Researcher → Explanation runner: configures run
- Explanation runner → Experiment configuration: uses settings
run.py - Explanation runner → Dataset loaders: loads data
run.py - Explanation runner → Model trainer: trains model
run.py - Model trainer → Dataset loaders: consumes loaders
modeltrainer.py - Model trainer → Model wrappers: wraps models
modeltrainer.py - Model trainer → Sequence models: selects model
modeltrainer.py - Model trainer → CNN models: selects model
modeltrainer.py - Model trainer → Transformer models: selects model
modeltrainer.py - Transformer models → Interpretability layers: uses layers
models/transformer.py - Explanation runner → Explainer base: sets up explainer
run.py - Explanation runner → FIT explainer: configures FIT
run.py - Explanation runner → WinIT explainer: configures WinIT
run.py - Explanation runner → TimeX explainer: configures TimeX
run.py - Explanation runner → TimeX++ explainer: configures TimeX++
run.py - Explanation runner → SWING explainer: runs SWING
- Explainer base → Model wrappers: explains model
explainer/base.py - Explainer base → Feature generators: trains generators
explainer/base.py - FIT explainer → Explainer base: extends base
explainer/fit.py - FIT explainer → Feature generators: builds generator
explainer/fit.py - Explanation runner → Attribution visualization: visualizes results
run.py - Explanation runner → Masking utilities: uses masking
run.py