Skip to content

Home

Effector is a python package for global and regional feature effects.

How to use

You need a dataset

Must be a np.ndarray with shape (n_samples, n_features).

X = np.random.uniform(-1, 1, (100, 2))

plt.scatter(X[:, 0], X[:, 1], c="red", marker="x")
plt.xlabel("Feature 1")
plt.ylabel("Feature 2")
plt.xlim(-1.5, 1.5)
plt.ylim(-1.5, 1.5)
plt.show()

Dataset

And a model

The black-box model you want to interpret.

Must be a callable with signature model(X: np.ndarray[n_samples, n_features]) -> np.ndarray[n_samples].

def predict(x):
    y = np.zeros(x.shape[0])
    ind = x[:, 1] > 0
    y[ind] = 10*x[ind, 0]
    y[~ind] = -10*x[~ind, 0]
    return y + np.random.normal(0, 1, x.shape[0])

plt.scatter(X[:, 0], X[:, 1], c=predict(X), cmap="coolwarm", marker="x")
plt.xlabel("Feature 1")
plt.ylabel("Feature 2")
plt.xlim(-1.5, 1.5)
plt.ylim(-1.5, 1.5)
plt.colorbar()
plt.show()

Model

\[f(x) = \begin{cases} 10x_1 + \epsilon, & \text{if } x_2 > 0 \\ -10x_1 + \epsilon, & \text{otherwise} \end{cases}\]
Global Effect

As you do not know the model's form, you want to understand how Feature 1 relates to the model's output. You can use Effector:

effector.PDP(data=X, model=predict).plot(feature=0, heterogeneity="ice")

Global Effect

Regional Effect

The plot implies that the global effect of Feature 1 on the model's output is zero, but there is heterogeneity in the effect. You want to further investigate the sources of heterogeneity:

reg_pdp = effector.RegionalPDP(data=X, model=predict)
reg_pdp.show_partitioning(feature=0)

reg_pdp.plot(feature=0, node_idx=0, heterogeneity="ice", y_limits=(-13, 13))
reg_pdp.plot(feature=0, node_idx=1, heterogeneity="ice", y_limits=(-13, 13))
reg_pdp.plot(feature=0, node_idx=2, heterogeneity="ice", y_limits=(-13, 13))

Feature 0 - Full partition tree:
    Node id: 0, name: x_0, heter: 5.57 || nof_instances:   100 || weight: 1.00
            Node id: 1, name: x_0 | x_1 <= 0.04, heter: 2.78 || nof_instances:    50 || weight: 0.50
            Node id: 2, name: x_0 | x_1  > 0.04, heter: 1.03 || nof_instances:    50 || weight: 0.50
--------------------------------------------------
Feature 0 - Statistics per tree level:
    Level 0, heter: 5.57
            Level 1, heter: 1.90 || heter drop: 3.67 (65.85%)

Regional Effect

Regional Effect

Dive in

That was a simple example. Effector has many more methods for obtaining global and regional effects and for each method many parameters to customize your analysis: For a deeper dive, check out:

Methods

Effector implements the following methods:

Method Global Effect Regional Effect Paper
PDP PDP RegionalPDP PDP, ICE, GAGDET-PD
d-PDP DerPDP RegionalDerPDP d-PDP, d-ICE
ALE ALE RegionalALE ALE, GAGDET-ALE
RHALE RHALE RegionalRHALE RHALE, DALE
SHAP-DP ShapDP RegionalShapDP SHAP, GAGDET-DP