effector

Effector

effector an eXplainable AI package for tabular data. It:

• creates global and regional effect plots
• has a simple API with smart defaults, but can become flexible if needed
• is model agnostic; can explain any underlying ML model
• integrates easily with popular ML libraries, like Scikit-Learn, Tensorflow and Pytorch
• is fast, for both global and regional methods
• provides a large collection of global and regional effects methods

📖 Documentation | 🔍 Intro to global and regional effects | 🔧 API | 🏗 Examples

Installation

Effector requires Python 3.10+:

pip install effector

Dependencies: numpy, scipy, matplotlib, tqdm, shap.

Quickstart

Train an ML model

import effector
import keras
import numpy as np
import tensorflow as tf

np.random.seed(42)
tf.random.set_seed(42)

# Load dataset
bike_sharing = effector.datasets.BikeSharing(pcg_train=0.8)
X_train, Y_train = bike_sharing.x_train, bike_sharing.y_train
X_test, Y_test = bike_sharing.x_test, bike_sharing.y_test

# Define and train a neural network
model = keras.Sequential([
    keras.layers.Dense(1024, activation="relu"),
    keras.layers.Dense(512, activation="relu"),
    keras.layers.Dense(256, activation="relu"),
    keras.layers.Dense(1)
])
model.compile(optimizer="adam", loss="mse", metrics=["mae", keras.metrics.RootMeanSquaredError()])
model.fit(X_train, Y_train, batch_size=512, epochs=20, verbose=1)
model.evaluate(X_test, Y_test, verbose=1)

Wrap it in a callable

def predict(x):
    return model(x).numpy().squeeze()

Explain it with global effect plots

# Initialize the Partial Dependence Plot (PDP) object
pdp = effector.PDP(
    X_test,  # Use the test set as background data
    predict,  # Prediction function
    feature_names=bike_sharing.feature_names,  # (optional) Feature names
    target_name=bike_sharing.target_name  # (optional) Target variable name
)

# Plot the effect of a feature
pdp.plot(
    feature=3,  # Select the 3rd feature (feature: hour)
    nof_ice=200,  # (optional) Number of Individual Conditional Expectation (ICE) curves to plot
    scale_x={"mean": bike_sharing.x_test_mu[3], "std": bike_sharing.x_test_std[3]},  # (optional) Scale x-axis
    scale_y={"mean": bike_sharing.y_test_mu, "std": bike_sharing.y_test_std},  # (optional) Scale y-axis
    centering=True,  # (optional) Center PDP and ICE curves
    show_avg_output=True,  # (optional) Display the average prediction
    y_limits=[-200, 1000]  # (optional) Set y-axis limits
)

Explain it with regional effect plots

# Initialize the Regional Partial Dependence Plot (RegionalPDP)
r_pdp = effector.RegionalPDP(
    X_test,  # Test set data
    predict,  # Prediction function
    feature_names=bike_sharing.feature_names,  # Feature names
    target_name=bike_sharing.target_name  # Target variable name
)

# Summarize the subregions of the 3rd feature (temperature)
r_pdp.summary(
    features=3,  # Select the 3rd feature for the summary
    scale_x_list=[  # scale each feature with mean and std
        {"mean": bike_sharing.x_test_mu[i], "std": bike_sharing.x_test_std[i]}
        for i in range(X_test.shape[1])
    ]
)

Feature 3 - Full partition tree:
🌳 Full Tree Structure:
───────────────────────
hr 🔹 [id: 0 | heter: 0.43 | inst: 3476 | w: 1.00]
    workingday = 0.00 🔹 [id: 1 | heter: 0.36 | inst: 1129 | w: 0.32]
        temp ≤ 6.50 🔹 [id: 3 | heter: 0.17 | inst: 568 | w: 0.16]
        temp > 6.50 🔹 [id: 4 | heter: 0.21 | inst: 561 | w: 0.16]
    workingday ≠ 0.00 🔹 [id: 2 | heter: 0.28 | inst: 2347 | w: 0.68]
        temp ≤ 6.50 🔹 [id: 5 | heter: 0.19 | inst: 953 | w: 0.27]
        temp > 6.50 🔹 [id: 6 | heter: 0.20 | inst: 1394 | w: 0.40]
--------------------------------------------------
Feature 3 - Statistics per tree level:
🌳 Tree Summary:
─────────────────
Level 0🔹heter: 0.43
    Level 1🔹heter: 0.31 | 🔻0.12 (28.15%)
        Level 2🔹heter: 0.19 | 🔻0.11 (37.10%)

The summary of feature hr (hour) says that its effect on the output is highly dependent on the value of features:

• workingday, wheteher it is a workingday or not
• temp, what is the temperature the specific hour

Let's see how the effect changes on these subregions!

Is it workingday or not?

# Plot regional effects after the first-level split (workingday vs non-workingday)
for node_idx in [1, 2]:  # Iterate over the nodes of the first-level split
    r_pdp.plot(
        feature=3,  # Feature 3 (temperature)
        node_idx=node_idx,  # Node index (1: workingday, 2: non-workingday)
        nof_ice=200,  # Number of ICE curves
        scale_x_list=[  # Scale features by mean and std
            {"mean": bike_sharing.x_test_mu[i], "std": bike_sharing.x_test_std[i]}
            for i in range(X_test.shape[1])
        ],
        scale_y={"mean": bike_sharing.y_test_mu, "std": bike_sharing.y_test_std},  # Scale the target
        y_limits=[-200, 1000]  # Set y-axis limits
    )

Is it hot or cold?

# Plot regional effects after second-level splits (workingday vs non-workingday and hot vs cold temperature)
for node_idx in [3, 4, 5, 6]:  # Iterate over the nodes of the second-level splits
    r_pdp.plot(
        feature=3,  # Feature 3 (temperature)
        node_idx=node_idx,  # Node index (hot/cold temperature and workingday/non-workingday)
        nof_ice=200,  # Number of ICE curves
        scale_x_list=[  # Scale features by mean and std
            {"mean": bike_sharing.x_test_mu[i], "std": bike_sharing.x_test_std[i]}
            for i in range(X_test.shape[1])
        ],
        scale_y={"mean": bike_sharing.y_test_mu, "std": bike_sharing.y_test_std},  # Scale target
        y_limits=[-200, 1000]  # Set y-axis limits
    )

Supported Methods

effector implements global and regional effect methods:

Method	Global Effect	Regional Effect	Reference	ML model	Speed
PDP	PDP	RegionalPDP	PDP	any	Fast for a small dataset
d-PDP	DerPDP	RegionalDerPDP	d-PDP	differentiable	Fast for a small dataset
ALE	ALE	RegionalALE	ALE	any	Fast
RHALE	RHALE	RegionalRHALE	RHALE	differentiable	Very fast
SHAP-DP	ShapDP	RegionalShapDP	SHAP	any	Fast for a small dataset and a light ML model

Method Selection Guide

From the runtime persepective there are three criterias:

• is the dataset small (N<10K) or large (N>10K instances) ?
• is the ML model light (runtime < 0.1s) or heavy (runtime > 0.1s) ?
• is the ML model differentiable or non-differentiable ?

Trust us and follow this guide:

• light + small + differentiable = any([PDP, RHALE, ShapDP, ALE, DerPDP])
• light + small + non-differentiable: [PDP, ALE, ShapDP]
• heavy + small + differentiable = any([PDP, RHALE, ALE, DerPDP])
• heavy + small + non differentiable = any([PDP, ALE])
• big + not differentiable = ALE
• big + differentiable = RHALE

Citation

If you use effector, please cite it:

@misc{gkolemis2024effector,
  title={effector: A Python package for regional explanations},
  author={Vasilis Gkolemis et al.},
  year={2024},
  eprint={2404.02629},
  archivePrefix={arXiv},
  primaryClass={cs.LG}
}

Spotlight on effector

📚 Featured Publications

• Gkolemis, Vasilis, et al.
  "Fast and Accurate Regional Effect Plots for Automated Tabular Data Analysis."
  Proceedings of the VLDB Endowment | ISSN 2150-8097

🎤 Talks & Presentations

• LMU-IML Group Talk
  Slides & Materials | LMU-IML Research
• AIDAPT Plenary Meeting
  Deep dive into effector
• XAI World Conference 2024
  Poster | Paper

🌍 Adoption & Collaborations

• AIDAPT Project
  Leveraging effector for explainable AI solutions.

🔍 Additional Resources

• Medium Post
  Effector: An eXplainability Library for Global and Regional Effects

• Courses & Lists:
  IML Course @ LMU
  Awesome ML Interpretability
  Awesome XAI
  Best of ML Python

📚 Related Publications

Papers that have inspired effector:

• REPID: Regional Effects in Predictive Models
  Herbinger et al., 2022 - Link

• Decomposing Global Feature Effects Based on Feature Interactions
  Herbinger et al., 2023 - Link

• RHALE: Robust Heterogeneity-Aware Effects
  Gkolemis Vasilis et al., 2023 - Link

• DALE: Decomposing Global Feature Effects
  Gkolemis Vasilis et al., 2023 - Link

• Greedy Function Approximation: A Gradient Boosting Machine
  Friedman, 2001 - Link

• Visualizing Predictor Effects in Black-Box Models
  Apley, 2016 - Link

• SHAP: A Unified Approach to Model Interpretation
  Lundberg & Lee, 2017 - Link

• Regionally Additive Models: Explainable-by-design models minimizing feature interactions
  Gkolemis Vasilis et al., 2023 - Link

License

effector is released under the MIT License.