Add TFTExplainer

CHANGELOG.md

@@ -23,10 +23,11 @@ but cannot always guarantee backwards compatibility. Changes that may **break co
   - `RegressionEnsembleModel` and `NaiveEnsembleModel` can generate probabilistic forecasts, probabilistics `forecasting_models` can be sampled to train the `regression_model`, updated the documentation (stacking technique). [#1692](https://github.com/unit8co/darts/pull/#1692) by [Antoine Madrona](https://github.com/madtoinou).
 - Improvements to `ShapExplainer`:
   - Added static covariates support to `ShapeExplainer`. [#1803](https://github.com/unit8co/darts/pull/#1803) by [Anne de Vries](https://github.com/anne-devries) and [Dennis Bader](https://github.com/dennisbader).
-  - Improved static covariates column naming when applying a `sklearn.preprocessing.OneHotEncoder` with `StaticCovariatesTransformer` [#1863](https://github.com/unit8co/darts/pull/1863) by [Anne de Vries](https://github.com/anne-devries)
-- Added `MSTL` (Season-Trend decomposition using LOESS for multiple seasonalities) as a `method` option for `extract_trend_and_seasonality()`. [#1879](https://github.com/unit8co/darts/pull/1879) by [Alex Colpitts](https://github.com/alexcolpitts96)
-- Added `RINorm` (Reversible Instance Norm) as a new layer normalization option. [#1121](https://github.com/unit8co/darts/issues/1121) by [Alex Colpitts](https://github.com/alexcolpitts96)
+  - Improved static covariates column naming when applying a `sklearn.preprocessing.OneHotEncoder` with `StaticCovariatesTransformer` [#1863](https://github.com/unit8co/darts/pull/1863) by [Anne de Vries](https://github.com/anne-devries).
+- Added `MSTL` (Season-Trend decomposition using LOESS for multiple seasonalities) as a `method` option for `extract_trend_and_seasonality()`. [#1879](https://github.com/unit8co/darts/pull/1879) by [Alex Colpitts](https://github.com/alexcolpitts96).
+- Added `RINorm` (Reversible Instance Norm) as a new layer normalization option. [#1121](https://github.com/unit8co/darts/issues/1121) by [Alex Colpitts](https://github.com/alexcolpitts96).
 - New forecasting model: `TiDEModel`  as proposed in [this paper](https://arxiv.org/abs/2304.08424). An MLP based encoder-decoder model that outperforms many Transformer-based architectures. [#1727](https://github.com/unit8co/darts/pull/1727) by [Alex Colpitts](https://github.com/alexcolpitts96).
+- New forecasting model explainer: `TFTExplainer` for `TFTModel`. You can now access and visualize the trained model's feature importances and self attention. [#1392](https://github.com/unit8co/darts/issues/1392) by [Sebastian Cattes](https://github.com/Cattes) and [Dennis Bader](https://github.com/dennisbader).
 
 **Fixed**
 - Fixed an issue not considering original component names for `TimeSeries.plot()` when providing a label prefix. [#1783](https://github.com/unit8co/darts/pull/1783) by [Simon Sudrich](https://github.com/sudrich).
@@ -36,10 +37,16 @@ but cannot always guarantee backwards compatibility. Changes that may **break co
 - Fixed `TimeSeries.__getitem__()` for series with a RangeIndex with start != 0 and freq != 1. [#1868](https://github.com/unit8co/darts/pull/#1868) by [Dennis Bader](https://github.com/dennisbader).
 - Fixed an issue where `DTWAlignment.plot_alignment()` was not plotting the alignment plot of series with a RangeIndex correctly. [#1880](https://github.com/unit8co/darts/pull/1880) by [Ahmet Zamanis](https://github.com/AhmetZamanis) and [Dennis Bader](https://github.com/dennisbader).
 - Fixed an issue when calling `ARIMA.predict()` and `num_samples > 1` (probabilistic forecasting), where the start point of the simulation was not anchored to the end of the target series. [#1893](https://github.com/unit8co/darts/pull/1893) by [Dennis Bader](https://github.com/dennisbader).
+- Fixed an issue when using `TFTModel.predict()` with `full_attention=True` where the attention mask was not applied properly. [#1392](https://github.com/unit8co/darts/issues/1392) by [Dennis Bader](https://github.com/dennisbader).
 
 **Removed**
 - Removed support for Python 3.7 [#1864](https://github.com/unit8co/darts/pull/#1864) by [Dennis Bader](https://github.com/dennisbader).
 
+### For developers of the library:
+
+**Improvements**
+- Refactored the `ForecastingModelExplainer` and `ExplainabilityResult` to simplify implementation of new explainers. [#1392](https://github.com/unit8co/darts/issues/1392) by [Dennis Bader](https://github.com/dennisbader). 
+
 ## [0.24.0](https://github.com/unit8co/darts/tree/0.24.0) (2023-04-12)
 ### For users of the library:
 
@@ -95,6 +102,8 @@ but cannot always guarantee backwards compatibility. Changes that may **break co
 - Major refactor of data transformers which simplifies implementation of new transformers. [#1409](https://github.com/unit8co/darts/pull/1409) by [Matt Bilton](https://github.com/mabilton).
 
 
+- Added new `TFTExplainer` class to implement the Explainable AI part described in [the paper](https://arxiv.org/abs/1912.09363) of the `TFT` model. [#1392](https://github.com/unit8co/darts/pull/1392) by [Sebastian Cattes](https://github.com/cattes).
+
 ## [0.23.1](https://github.com/unit8co/darts/tree/0.23.1) (2023-01-12)
 Patch release
 
@@ -166,7 +175,6 @@ Patch release
   by [Antoine Madrona](https://github.com/madtoinou).
 
 
-
 **Fixed**
 - Fixed edge case in ShapExplainer for regression models where covariates series > target series
   [#1310](https://https://github.com/unit8co/darts/pull/1310) by [Rijk van der Meulen](https://github.com/rijkvandermeulen)

darts/explainability/__init__.py

@@ -3,5 +3,10 @@
 --------------
 """
 
-from darts.explainability.explainability_result import ExplainabilityResult
+from darts.explainability.explainability_result import (
+    ShapExplainabilityResult,
+    TFTExplainabilityResult,
+    _ExplainabilityResult,
+)
 from darts.explainability.shap_explainer import ShapExplainer
+from darts.explainability.tft_explainer import TFTExplainer