book

Interpretable Machine Learning with Python - Second Edition

by Serg Masís

October 2023

Intermediate to advanced

606 pages

16h 37m

English

Packt Publishing

Read now

Unlock full access

Who this book is forWhat this book coversTo get the most out of this bookGet in touch
Technical requirementsWhat is machine learning interpretation?Understanding a simple weight prediction modelUnderstanding the difference between interpretability and explainabilityWhat is interpretability?Beware of complexityWhen does interpretability matter?What are black-box models?What are white-box models?What is explainability?Why and when does explainability matter?A business case for interpretabilityBetter decisionsMore trusted brandsMore ethicalMore profitableSummaryImage sourcesDataset sourcesFurther reading
Technical requirementsThe missionDetails about CVDThe approachPreparationsLoading the librariesUnderstanding and preparing the dataThe data dictionaryData preparationInterpretation method types and scopesModel interpretability method typesModel interpretability scopesInterpreting individual predictions with logistic regressionAppreciating what hinders machine learning interpretabilityNon-linearityInteractivityNon-monotonicityMission accomplishedSummaryFurther reading
Technical requirementsThe missionThe approachThe preparationsLoading the librariesUnderstanding and preparing the dataThe data dictionaryData preparationReviewing traditional model interpretation methodsPredicting minutes delayed with various regression methodsClassifying flights as delayed or not delayed with various classification methodsTraining and evaluating the classification modelsUnderstanding limitations of traditional model interpretation methodsStudying intrinsically interpretable (white-box) modelsGeneralized Linear Models (GLMs)Linear regressionRidge regressionPolynomial regressionLogistic regressionDecision treesCART decision treesRuleFitInterpretation and feature importanceNearest neighborsk-Nearest NeighborsNaïve BayesGaussian Naïve BayesRecognizing the trade-off between performance and interpretabilitySpecial model propertiesThe key property: explainabilityThe remedial property: regularizationAssessing performanceDiscovering newer interpretable (glass-box) modelsExplainable Boosting Machine (EBM)Global interpretationLocal interpretationPerformanceGAMI-NetGlobal interpretationLocal interpretationPerformanceMission accomplishedSummaryDataset sourcesFurther reading
Technical requirementsThe missionThe approachThe preparationsLoading the librariesData preparationModel training and evaluationWhat is feature importance?Assessing feature importance with model-agnostic methodsPermutation feature importanceSHAP valuesComprehensive explanations with KernelExplainerFaster explanations with TreeExplainerVisualize global explanationsSHAP bar plotSHAP beeswarm plotFeature summary explanationsPartial dependence plots SHAP scatter plotALE plotsFeature interactionsSHAP bar plot with clustering2D ALE plotsPDP interactions plotsMission accomplishedSummaryFurther reading
Technical requirementsThe missionThe approachThe preparationsLoading the librariesUnderstanding and preparing the dataThe data dictionaryData preparationLeveraging SHAP’s KernelExplainer for local interpretations with SHAP valuesTraining a C-SVC modelComputing SHAP values using KernelExplainerLocal interpretation for a group of predictions using decision plotsLocal interpretation for a single prediction at a time using a force plotEmploying LIMEWhat is LIME?Local interpretation for a single prediction at a time using LimeTabularExplainerUsing LIME for NLPTraining a LightGBM modelLocal interpretation for a single prediction at a time using LimeTextExplainerTrying SHAP for NLPComparing SHAP with LIMEMission accomplishedSummaryDataset sourcesFurther reading
Technical requirementsThe missionUnfair bias in recidivism risk assessmentsThe approachThe preparationsLoading the librariesUnderstanding and preparing the dataThe data dictionaryExamining predictive bias with confusion matricesData preparationModelingGetting acquainted with our “instance of interest”Understanding anchor explanationsPreparations for anchor and counterfactual explanations with alibiLocal interpretations for anchor explanationsExploring counterfactual explanationsCounterfactual explanations guided by prototypesCounterfactual instances and much more with WITConfiguring WITDatapoint editorPerformance & FairnessMission accomplishedSummaryDataset sourcesFurther reading
Technical requirementsThe missionThe approachPreparationsLoading the librariesUnderstanding and preparing the dataData preparationInspect dataThe CNN modelsLoad the CNN modelAssessing the CNN classifier with traditional interpretation methodsDetermining what misclassifications to focus onVisualizing the learning process with activation-based methodsIntermediate activationsEvaluating misclassifications with gradient-based attribution methodsSaliency mapsGuided Grad-CAMIntegrated gradientsBonus method: DeepLIFTTying it all togetherUnderstanding classifications with perturbation-based attribution methodsFeature ablationOcclusion sensitivityShapley value samplingKernelSHAPTying it all togetherMission accomplishedSummaryFurther reading
Technical requirementsThe missionThe approachThe preparationsLoading the librariesUnderstanding and preparing the dataThe data dictionaryLoading the modelVisualizing attention with BertVizPlotting all attention with the model viewDiving into layer attention with the head viewInterpreting token attributions with integrated gradientsLIME, counterfactuals, and other possibilities with the LITMission accomplishedSummaryFurther reading
Technical requirementsThe missionThe approachThe preparationLoading the librariesUnderstanding and preparing the dataThe data dictionaryUnderstanding the dataData preparationLoading the LSTM modelAssessing time series models with traditional interpretation methodsUsing standard regression metricsPredictive error aggregationsEvaluating the model like a classification problemGenerating LSTM attributions with integrated gradientsComputing global and local attributions with SHAP’s KernelExplainerWhy use KernelExplainer?Defining a strategy to get it to work with a multivariate time series modelLaying the groundwork for the permutation approximation strategyComputing the SHAP valuesIdentifying influential features with factor prioritizationComputing Morris sensitivity indicesAnalyzing the elementary effectsQuantifying uncertainty and cost sensitivity with factor fixingGenerating and predicting on Saltelli samplesPerforming Sobol sensitivity analysisIncorporating a realistic cost functionMission accomplishedSummaryDataset and image sourcesFurther reading

Technical requirementsThe missionThe approachThe preparationsLoading the librariesUnderstanding and preparing the dataUnderstanding the effect of irrelevant featuresCreating a base modelEvaluating the modelTraining the base model at different max depthsReviewing filter-based feature selection methodsBasic filter-based methodsConstant features with a variance thresholdQuasi-constant features with value_countsDuplicating featuresRemoving unnecessary featuresCorrelation filter-based methodsRanking filter-based methodsComparing filter-based methodsExploring embedded feature selection methodsDiscovering wrapper, hybrid, and advanced feature selection methodsWrapper methodsSequential forward selection (SFS)Hybrid methodsRecursive Feature Elimination (RFE)Advanced methodsModel-agnostic feature importanceGenetic algorithmsEvaluating all feature-selected modelsConsidering feature engineeringMission accomplishedSummaryDataset sourcesFurther reading
Technical requirementsThe missionThe approachThe preparationsLoading the librariesUnderstanding and preparing the dataThe data dictionaryData preparationDetecting biasVisualizing dataset biasQuantifying dataset biasQuantifying model biasMitigating biasPreprocessing bias mitigation methodsThe Reweighing methodThe disparate impact remover methodIn-processing bias mitigation methodsThe exponentiated gradient reduction methodThe gerry fair classifier methodPost-processing bias mitigation methodsThe equalized odds post-processing methodThe calibrated equalized odds postprocessing methodTying it all together!Creating a causal modelUnderstanding the results of the experimentUnderstanding causal modelsInitializing the linear doubly robust learnerFitting the causal modelUnderstanding heterogeneous treatment effectsChoosing policiesTesting estimate robustnessAdding a random common causeReplacing the treatment variable with a random variableMission accomplishedSummaryDataset sourcesFurther reading
Technical requirementsThe missionThe approachThe preparationsLoading the librariesUnderstanding and preparing the dataVerifying the sampling balancePlacing guardrails with feature engineeringOrdinalizationDiscretizationInteraction terms and non-linear transformationsCategorical encodingOther preparationsTuning models for interpretabilityTuning a Keras neural networkDefining the model and parameters to tuneRunning the hyperparameter tuningExamining the resultsEvaluating the best modelTuning other popular model classesA quick introduction to relevant model parametersBatch hyperparameter tuning modelsEvaluating models by precisionAssessing fairness for the highest-performing modelOptimizing for fairness with Bayesian hyperparameter tuning and custom metricsDesigning a custom metricRunning Bayesian hyperparameter tuningFitting and evaluating a model with the best parametersExamining racial bias through feature importanceImplementing model constraintsConstraints for XGBoostSetting regularization and constraint parametersTraining and evaluating the constrained modelExamining constraintsConstraints for TensorFlow LatticeInitializing the model and Lattice inputsBuilding a Keras model with TensorFlow Lattice layersTraining and evaluating the modelMission accomplishedSummaryDataset sourcesFurther reading
Technical requirementsThe missionThe approachThe preparationsLoading the librariesUnderstanding and preparing the dataLoading the CNN base modelAssessing the CNN base classifierLearning about evasion attacksFast gradient sign method attackCarlini and Wagner infinity norm attackTargeted adversarial patch attackDefending against targeted attacks with preprocessingShielding against any evasion attack by adversarial training of a robust classifierEvaluating adversarial robustnessComparing model robustness with attack strengthMission accomplishedSummaryDataset sourcesFurther reading
Understanding the current landscape of ML interpretabilityTying everything together!Current trendsSpeculating on the future of ML interpretabilityA new vision for MLA multidisciplinary approachAdequate standardizationEnforcing regulationSeamless machine learning automation with built-in interpretationTighter integration with MLOps engineersSummaryFurther reading

Content preview from Interpretable Machine Learning with Python - Second Edition

8 Interpreting NLP Transformers

In the last chapter, we learned about applying explanation methods to a specific type of deep learning model architecture, convolutional neural networks. In this chapter, we will provide some tools to do the same with the transformer model architecture. Transformer models are becoming increasingly popular, and their most common use case is Natural Language Processing (NLP). We broached the subject of NLP in Chapter 5, Local Model-Agnostic Interpretation Methods. In this chapter, we will do so too but with transformer-specific methods and tools. First, we will discuss how to visualize attention mechanisms, followed by interpreting integrated gradient attributions, and lastly, exploring the Swiss Army knife that ...