Explainable Al (XAI) with Python

Demonstrate glass box models for explainable ai in Python by loading, preprocessing, and exploring the stroke dataset. Implement logistic regression and interpretable models, and address imbalanced data and encoding techniques.

Demonstrates preprocessing for a health care dataset in explainable AI with Python, including one-hot encoding of categorical features, handling missing values, dropping id, and 80/20 train-test split for stroke prediction.

Master the need for train-test split in explainable AI with Python by dividing data into training and testing sets to prevent overfitting and accurately evaluate model performance.

Balance the imbalanced stroke dataset by exploring under sampling, oversampling, and synthetic data with SMOTE, and implement random oversampling to equalize class distribution.

Balance an unbalanced dataset by applying random oversampling to the minority class after preprocessing steps like encoding the categorical attribute, imputing missing bmi with zero, and dropping the id column.

Explore the confusion matrix to evaluate classifier performance, defining true positives, true negatives, false positives, and false negatives, then compute precision, recall, and the F1 score.

Explore the interpretML package to explain glass box and black box models, delivering global and local explanations on the stroke dataset using logistic regression and other classifiers.

Explore the explainable boosting classifier, a glass-box, tree-based generalized additive model that delivers linear regression-like interpretability with automatic interaction predictions and accuracy comparable to black-box models.

Learn how lime provides local, interpretable, model-agnostic explanations for individual predictions of any black-box model using a simple surrogate.

Explore how lime constructs a local, simple surrogate around a specific input to explain a complex loan-approval model, using linear regression or a simple decision tree within a neighborhood.

LIME builds a local, sparse linear model around a data point by perturbing features, weighting nearby points, and mimicking a complex model’s predictions for explanation.

Explore how LIME tabular explains a random forest on a balanced stroke dataset, generating 20 local, instance-level explanations and visual insights for explainable AI with Python.

Demonstrates lime, a local, model-agnostic explainer, for a two-class text dataset using Python in Google Colab. Shows tf-idf and bag-of-words preprocessing on newsgroup data and lime explanations via notebooks.

Demonstrate lime explanations for textual data by evaluating five classifiers on the newsgroup datasets, highlighting the best model and generating local, model-agnostic explanations with keyword highlights.

Explore the mathematical modelling of Shapley values for explainable AI by decoding the SHAP equation, calculating feature contributions (age, BMI) from a stroke dataset, and understanding subset effects.

Explore explainable ai with python by showing how Shapley values allocate the contribution of age, gender, and job in a black-box income predictor through all feature coalitions and marginal contributions.

Explore how SHAP values are modeled through marginal contributions, binomial coefficients, and reciprocal edge weights, and derive feature attributions for a full model versus a null model.

Counterfactual explanations reveal the smallest changes to input features needed to flip a model’s prediction without exposing the black box, such as lowering body mass index to avoid a stroke.

Explore mathematical modeling of counterfactual explanations in explainable ai with Python, minimally altering input features to flip predictions, comparing brute-force and model-aware approaches, and emphasizing visible counterfactuals and cost metrics.

Explore global counterfactuals to reveal the model's overall behavior and detect bias across racial and gender subgroups, useful for regulators assessing fairness in loan approvals.

Demonstrate the what-if tool by visualizing and generating counterfactual explanations, exploring model behavior with preloaded census models, data point edits, and partial dependence plots.

Demonstrates Google's what-if tool to analyze fairness in recidivism classification, exploring data points, counterfactuals, and partial dependence plots to reveal model bias and perform fairness and performance analysis.

Explore explainable ai with python through interactive lrp demos, generating real-time heatmaps that justify handwriting, image, and text classifications and visual question answering decisions.

Explore how layer-wise relevance propagation explains neural network predictions by propagating relevance backward through CNN layers to produce pixel-level heatmaps.

Model layer-wise relevance propagation (LRP) to compute per-pixel relevance scores and generate heatmaps that explain cnn image classifications, propagating backward from the predicted class through layers.

Explainable ai with python presents the contrastive explanations method (cem), highlighting minimal present and absent features, pertinent positives and negatives, and applications to neural networks, medicine, and image analysis.