Insights module and SHAPInsight

CHANGELOG.md

@@ -6,6 +6,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 ### Added
+- Added SHAP analysis within the new `diagnostics` package.
 - `allow_missing` and `allow_extra` keyword arguments to `Objective.transform`
 
 ### Deprecations

CONTRIBUTORS.md

@@ -25,4 +25,6 @@
 - Di Jin (Merck Life Science KGaA, Darmstadt, Germany):\
   Cardinality constraints
 - Julian Streibel (Merck Life Science KGaA, Darmstadt, Germany):\
-  Bernoulli multi-armed bandit and Thompson sampling
+  Bernoulli multi-armed bandit and Thompson sampling
+- Alexander Wieczorek (Swiss Federal Institute for Materials Science and Technology, Dübendorf, Switzerland):\
+  SHAP explainers for diagnoatics

README.md

@@ -296,6 +296,7 @@ The available groups are:
 - `mypy`: Required for static type checking.
 - `onnx`: Required for using custom surrogate models in [ONNX format](https://onnx.ai).
 - `polars`: Required for optimized search space construction via [Polars](https://docs.pola.rs/)
+- `diagnostics`: Required for feature importance ranking via [SHAP](https://shap.readthedocs.io/)
 - `simulation`: Enabling the [simulation](https://emdgroup.github.io/baybe/stable/_autosummary/baybe.simulation.html) module.
 - `test`: Required for running the tests.
 - `dev`: All of the above plus `tox` and `pip-audit`. For code contributors.

baybe/_optional/diagnostics.py

@@ -0,0 +1,16 @@
+"""Optional import for diagnostics utilities."""
+
+from baybe.exceptions import OptionalImportError
+
+try:
+    import shap
+except ModuleNotFoundError as ex:
+    raise OptionalImportError(
+        "Explainer functionality is unavailable because 'diagnostics' is not installed."
+        " Consider installing BayBE with 'diagnostics' dependency, e.g. via "
+        "`pip install baybe[diagnostics]`."
+    ) from ex
+
+__all__ = [
+    "shap",
+]

baybe/_optional/info.py

@@ -28,6 +28,7 @@ def exclude_sys_path(path: str, /):  # noqa: DOC402, DOC404
     MORDRED_INSTALLED = find_spec("mordred") is not None
     ONNX_INSTALLED = find_spec("onnxruntime") is not None
     POLARS_INSTALLED = find_spec("polars") is not None
+    SHAP_INSTALLED = find_spec("shap") is not None
     PRE_COMMIT_INSTALLED = find_spec("pre_commit") is not None
     PYDOCLINT_INSTALLED = find_spec("pydoclint") is not None
     RDKIT_INSTALLED = find_spec("rdkit") is not None
@@ -45,6 +46,7 @@ def exclude_sys_path(path: str, /):  # noqa: DOC402, DOC404
 
 # Package combinations
 CHEM_INSTALLED = MORDRED_INSTALLED and RDKIT_INSTALLED
+DIAGNOSTICS_INSTALLED = SHAP_INSTALLED
 LINT_INSTALLED = all(
     (
         FLAKE8_INSTALLED,

baybe/utils/diagnostics.py

@@ -0,0 +1,194 @@
+"""Diagnostics utilities."""
+
+import numbers
+import warnings
+
+import numpy as np
+import pandas as pd
+
+from baybe import Campaign
+from baybe._optional.diagnostics import shap
+from baybe.utils.dataframe import to_tensor
+
+
+def explainer(
+    campaign: Campaign,
+    explainer_class: shap.Explainer = shap.KernelExplainer,
+    computational_representation: bool = False,
+    **kwargs,
+) -> shap.Explainer:
+    """Create an explainer for the provided campaign.
+
+    Args:
+        campaign: The campaign to be explained.
+        explainer_class: The explainer to be used. Default is shap.KernelExplainer.
+        computational_representation: Whether to compute the Shapley values
+            in  computational or experimental searchspace.
+            Default is False.
+        **kwargs: Additional keyword arguments to be passed to the explainer.
+
+    Returns:
+        The explainer for the provided campaign.
+
+    Raises:
+        ValueError: If no measurements have been provided yet.
+    """
+    if campaign.measurements.empty:
+        raise ValueError("No measurements have been provided yet.")
+
+    data = campaign.measurements[[p.name for p in campaign.parameters]].copy()
+
+    if computational_representation:
+        data = campaign.searchspace.transform(data)
+
+        def model(x):
+            tensor = to_tensor(x)
+            output = campaign.get_surrogate()._posterior_comp(tensor).mean
+
+            return output.detach().numpy()
+    else:
+
+        def model(x):
+            df = pd.DataFrame(x, columns=data.columns)
+            output = campaign.get_surrogate().posterior(df).mean
+
+            return output.detach().numpy()
+
+    shap_explainer = explainer_class(model, data, **kwargs)
+    return shap_explainer
+
+
+def explanation(
+    campaign: Campaign,
+    data: np.ndarray = None,
+    explainer_class: shap.Explainer = shap.KernelExplainer,
+    computational_representation: bool = False,
+    **kwargs,
+) -> shap.Explanation:
+    """Compute the Shapley values for the provided campaign and data.
+
+    Args:
+        campaign: The campaign to be explained.
+        data: The data to be explained.
+            Default is None which uses the campaign's measurements.
+        explainer_class: The explainer to be used.
+            Default is shap.KernelExplainer.
+        computational_representation: Whether to compute the Shapley values
+            in computational or experimental searchspace.
+            Default is False.
+        **kwargs: Additional keyword arguments to be passed to the explainer.
+
+    Returns:
+        The Shapley values for the provided campaign.
+
+    Raises:
+        ValueError: If the provided data does not have the same amount of parameters
+            as previously provided to the explainer.
+    """
+    is_shap_explainer = not explainer_class.__module__.startswith(
+        "shap.explainers.other."
+    )
+
+    if not is_shap_explainer and not computational_representation:
+        raise ValueError(
+            "Experimental representation is not "
+            "supported for non-Kernel SHAP explainer."
+        )
+
+    explainer_obj = explainer(
+        campaign,
+        explainer_class=explainer_class,
+        computational_representation=computational_representation,
+        **kwargs,
+    )
+
+    if data is None:
+        if isinstance(explainer_obj.data, np.ndarray):
+            data = explainer_obj.data
+        else:
+            data = explainer_obj.data.data
+    elif computational_representation:
+        data = campaign.searchspace.transform(data)
+
+    if not is_shap_explainer:
+        """Return attributions for non-SHAP explainers."""
+        if explainer_class.__module__.endswith("maple"):
+            """Additional argument for maple to increase comparability to SHAP."""
+            attributions = explainer_obj.attributions(data, multiply_by_input=True)[0]
+        else:
+            attributions = explainer_obj.attributions(data)[0]
+        if computational_representation:
+            feature_names = campaign.searchspace.comp_rep_columns
+        else:
+            feature_names = campaign.searchspace.parameter_names
+        explanations = shap.Explanation(
+            values=attributions,
+            base_values=data,
+            data=data,
+        )
+        explanations.feature_names = list(feature_names)
+        return explanations
+
+    if data.shape[1] != explainer_obj.data.data.shape[1]:
+        raise ValueError(
+            "The provided data does not have the same amount "
+            "of parameters as the shap explainer background."
+        )
+    else:
+        shap_explanations = explainer_obj(data)[:, :, 0]
+
+    return shap_explanations
+
+
+def plot_beeswarm(explanation: shap.Explanation, **kwargs) -> None:
+    """Plot the Shapley values using a beeswarm plot."""
+    shap.plots.beeswarm(explanation, **kwargs)
+
+
+def plot_waterfall(explanation: shap.Explanation, **kwargs) -> None:
+    """Plot the Shapley values using a waterfall plot."""
+    shap.plots.waterfall(explanation, **kwargs)
+
+
+def plot_bar(explanation: shap.Explanation, **kwargs) -> None:
+    """Plot the Shapley values using a bar plot."""
+    shap.plots.bar(explanation, **kwargs)
+
+
+def plot_scatter(explanation: shap.Explanation | memoryview, **kwargs) -> None:
+    """Plot the Shapley values using a scatter plot while leaving out string values.
+
+    Args:
+        explanation: The Shapley values to be plotted.
+        **kwargs: Additional keyword arguments to be passed to the scatter plot.
+
+    Raises:
+        ValueError: If the provided explanation object does not match the
+            required types.
+    """
+    if isinstance(explanation, memoryview):
+        data = explanation.obj
+    elif isinstance(explanation, shap.Explanation):
+        data = explanation.data.data.obj
+    else:
+        raise ValueError("The provided explanation argument is not of a valid type.")
+
+    def is_not_numeric_column(col):
+        return np.array([not isinstance(v, numbers.Number) for v in col]).any()
+
+    if data.ndim == 1:
+        if is_not_numeric_column(data):
+            warnings.warn(
+                "Cannot plot scatter plot for the provided "
+                "explanation as it contains non-numeric values."
+            )
+        else:
+            shap.plots.scatter(explanation, **kwargs)
+    else:
+        number_enum = [i for i, x in enumerate(data[1]) if not isinstance(x, str)]
+        if len(number_enum) < len(explanation.feature_names):
+            warnings.warn(
+                "Cannot plot SHAP scatter plot for all "
+                "parameters as some contain non-numeric values."
+            )
+        shap.plots.scatter(explanation[:, number_enum], **kwargs)

mypy.ini

@@ -64,3 +64,6 @@ ignore_missing_imports = True
 
 [mypy-polars]
 ignore_missing_imports = True
+
+[mypy-shap.*]
+ignore_missing_imports = True

pyproject.toml

@@ -81,6 +81,7 @@ onnx = [
 
 dev = [
     "baybe[chem]",
+    "baybe[diagnostics]",
     "baybe[docs]",
     "baybe[examples]",
     "baybe[lint]",
@@ -94,6 +95,11 @@ dev = [
     "uv>=0.3.0", #  `uv lock` (for lockfiles) is stable since 0.3.0: https://github.com/astral-sh/uv/issues/2679#event-13950215962
 ]
 
+diagnostics = [
+    "shap>=0.46.0",
+    "lime>=0.2.0.1"
+]
+
 docs = [
     "baybe[examples]", # docs cannot be built without running examples
     "furo>=2023.09.10",

pytest.ini

@@ -10,6 +10,8 @@ addopts =
     --ignore=baybe/_optional
     --ignore=baybe/utils/chemistry.py
     --ignore=tests/simulate_telemetry.py
+    --ignore=baybe/utils/diagnostics.py
+    --ignore=tests/utils/test_diagnostics.py
 testpaths = 
     baybe
     tests