Learning AutoML. Automating ML Pipelines with AutoGluon, Leading Frameworks, and Real-World Integration Kerem Tomak

Dostępność produktu

Produkt nie został jeszcze oceniony pod kątem ułatwień dostępu lub nie podano żadnych informacji o ułatwieniach dostępu lub są one niewystarczające. Prawdopodobnie Wydawca/Dostawca jeszcze nie umożliwił dokonania walidacji produktu lub nie przekazał odpowiednich informacji na temat jego dostępności.

Spis treści książki

• Foreword by Thomas H. Davenport
• Foreword by Gregory Wheeler
• Preface
  • Why I Wrote This Book
  • Who Should Read This Book
  • How This Book Is Organized
    • Part I: Foundations of AutoML
    • Part II: Core AutoML Techniques
    • Part III: AutoML for Different Data Types
    • Part IV: Production and MLOps
    • Part V: Case Studies
    • Reading Strategies for Different Goals
    • Hands-On Learning Approach
  • Conventions Used in This Book
  • Using Code Examples
  • OReilly Online Learning
  • How to Contact Us
  • Acknowledgments
• I. Foundations of AutoML
• 1. What Is Automated Machine Learning?
  • The Growing Demand for Machine Learning Solutions
  • Addressing the Data Science Talent Gap
  • Democratizing AI Development
  • AutoML in the Machine Learning Landscape
    • Open Source AutoML Libraries
    • Enterprise AutoML Platforms
    • Comparison of Leading Frameworks
  • Who Should Use AutoML?
  • AutoML Across Industries: Transforming Business Processes
    • Finance
    • Healthcare and Life Sciences
    • Retail and Ecommerce
    • Manufacturing
    • Other Sectors
    • The Tiered Use Case Model
  • Overcoming Hurdles: Persistent Challenges in AutoML
    • Interpretability (the Black Box Problem)
    • Need for Customization Versus Automation
    • Data Quality Dependency and Robustness
    • Computational Costs and Resource Intensity
    • Addressing Bias and Fairness
    • Scalability and Efficiency
  • The Horizon: Future Trends Shaping AutoML
    • Synergy with Large Language Models (LLMs) and Foundation Models
      • AutoML for LLMs
      • LLMs for AutoML
      • Foundation models as bases
      • Challenges
    • Next-Generation Neural Architecture Search (NAS)
    • Maturity of Multimodal Explainable AI (MXAI)
    • Continued Democratization via Low-Code/No-Code
    • Expansion to Edge Computing and Federated Learning
  • Summary
• 2. The Rise and Current State of AutoML
  • Early Automation (Pre-2010): Laying the Groundwork
    • Feature Selection
    • Hyperparameter Search
    • Meta-Learning Research
    • Limitations of Early Approaches
  • First Generation (20102015): Solving the CASH Problem
    • Auto-WEKA (2013)
    • Hyperopt (2013)
    • SMAC (Sequential Model-Based Algorithm Configuration)
    • First Generation Impact and Legacy
  • Second Generation (20152020): Solving the Usability and Enterprise Problem
    • Auto-sklearn (2015)
    • PyCaret (2020)
    • TPOT (Tree-Based Pipeline Optimization Tool) (2016)
    • H2O AutoML
    • Google Cloud AutoML (Now Part of Vertex AI)
    • Second Generation Impact and Legacy
  • Third Generation (2020Present): Solving the Multimodal and MLOps Problem
    • AutoGluon (Amazon)
    • Google Vertex AI
    • MLJAR and AWS SageMaker Autopilot
    • Third Generation Key Capabilities
      • Multimodal learning
      • Explainability integration (XAI)
      • MLOps integration
      • Neural architecture search (NAS)
    • The Emergence of LLM-Assisted AutoML
  • Summary
• 3. Understanding the AutoML Pipeline
  • The Architecture of Automated Machine Learning
  • Data Preprocessing
    • Data Quality Assessment and Cleaning
    • Missing Data Strategies
    • Data Validation and Integrity Checks
  • Feature Engineering
    • Multilevel Feature Generation
    • Domain-Specific Feature Engineering
    • Feature Selection and Pruning
    • Representation Learning Integration
  • Hyperparameter Optimization
    • Advanced Search Strategies
    • Multifidelity Optimization
    • Configuration Space Design
    • Budget-Aware Optimization
  • Neural Architecture Search
    • Search Space Engineering
    • Efficiency Techniques
    • Hardware-Aware Architecture Search
    • Architecture Transfer and Meta-Learning
  • Model Selection, Ensembling, and Stacking
    • Diversity-Driven Ensemble Construction
    • Advanced Stacking Techniques
    • Dynamic Ensemble Selection
    • Resource-Aware Ensemble Optimization
  • Model Deployment and Monitoring
    • Production Readiness Considerations
    • Scalability and Performance Optimization
    • Model Monitoring and Maintenance
    • Interpretability and Explainability
  • Pipeline Integration and Optimization
    • Cross-Stage Optimization Strategies
    • Resource Allocation and Management
    • Feedback Mechanisms and Continuous Learning
  • Challenges and Future Directions
    • Scalability and Efficiency
    • Robustness and Reliability
    • Democratization and Accessibility
    • The Double-Edged Sword of Democratization
  • Summary
• II. Core AutoML Techniques
• 4. Automated Data Preprocessing and Feature Engineering
  • Working Dataset: RetailMart Ecommerce Platform
  • Intelligent Data Profiling and Quality Assessment
  • Smart Data Type Handling and Transformation
    • DateTime Feature Extraction
    • Text Preprocessing Pipelines
  • Automated Feature Engineering
    • Traditional Feature Engineering Automation
    • Advanced Feature Learning Techniques
  • Intelligent Feature Selection and Dimensionality Management
  • Preprocessing Complex and Multimodal Data
  • Production-Ready Preprocessing Pipelines
  • Summary
• 5. Hyperparameter Optimization
  • The Challenge of Hyperparameter Optimization
    • The Computational Cost Challenge
    • The Sensitivity Problem
    • Real-World Impact
  • Grid Search Versus Random Search: Building the Foundation
    • Grid Search: Systematic but Limited
    • Random Search: A Surprisingly Effective Alternative
    • A Practical Comparison
    • Modern Implementation
    • When to Use Each Approach
    • Limitations of Both Approaches
  • Bayesian Optimization: Learning from Experience
    • The Core Insight
    • Surrogate Models
      • Gaussian processes (GPs)
      • Random forests
      • Tree-structured Parzen estimators (TPE)
    • Acquisition Functions
    • Real-World Success Stories
      • NLP model tuning
      • Reinforcement learning
      • Industrial forecasting
    • Modern Tools and Implementation
      • Optuna
      • Scikit-Optimize
      • Ray Tune
    • AWS SageMaker Automatic Model Tuning
    • Advanced Techniques
    • Practical Considerations
    • Limitations and Challenges
  • Early Stopping and Scheduling: Working Smarter, Not Harder
    • The Core Insight
    • Successive Halving: A Tournament Approach
    • Hyperband: Automating the Resource Allocation
    • Asynchronous Successive Halving (ASHA)
    • Population-Based Training: Evolution During Training
    • Layer Freezing: A Novel Fidelity Dimension
    • Practical Implementation
    • Real-World Results
    • Combining with Bayesian Optimization
    • When Early Stopping Works Best
    • A Cautionary Note
  • Multifidelity Optimization: Beyond Simple Early Stopping
    • The Multifidelity Paradigm
    • Advanced Multifidelity Methods
    • Practical Implementation
  • Case Study: Personal Investment Portfolio Optimization with Multifidelity HPO
    • Background and Problem Definition
    • Dataset and Features
    • Multifidelity Strategy Implementation
    • Implementation
    • Resource Management and Results
  • Key Insights and Practical Considerations
    • Early Stopping Effectiveness
    • Model Type Performance Patterns
    • Resource Allocation Strategy
    • Production Deployment Considerations
    • Lessons for Individual Practitioners
    • When to Use Multifidelity Optimization
  • Summary
• 6. Neural Architecture Search (NAS)
  • Understanding Neural Architecture Search
  • The Three Pillars of NAS
  • Search Space Design: Defining the Boundaries
    • The Art of Constraint
    • Types of Search Spaces
    • Task-Specific Considerations
    • Emerging Specialized Search Spaces
    • Multi-Objective Search Spaces
    • Balancing Efficiency and Discovery
    • The NAS-Bench Revolution
  • Search Strategies: Finding Needles in Haystacks
    • The Evolution of Search Strategies
    • Choosing Your Search Strategy: A Practical Perspective
    • Reinforcement Learning: The Original Approach
    • Evolutionary Algorithms: Nature-Inspired Search
    • Differentiable NAS: The Game Changer
    • Gradient-Based Methods and Advanced Techniques
    • Hybrid Approaches: Best of Both Worlds
    • Choosing the Right Strategy
  • Performance Estimation: The Efficiency Imperative
    • Bridging the Evaluation Gap
    • The Training Bottleneck
    • Multifidelity Evaluation: Training Less, Learning More
    • One-Shot Architecture Search: Train Once, Evaluate Many
    • Learning Curve Extrapolation
    • Zero-Cost Proxies: Instant Architecture Evaluation
    • Surrogate Models: Learning to Predict Performance
    • Combining Approaches for Maximum Efficiency
  • Efficient NAS: Making It Practical
    • The Efficiency Revolution
    • Building Production-Ready NAS Systems
    • Weight Sharing: The Foundation of Efficient NAS
    • Once-For-All Networks: Decoupling Training and Deployment
    • Progressive Search Strategies
    • Hardware-Aware Optimization
    • Production Deployment Tools for NAS-Discovered Architectures
      • AWS Deployment Stack
      • Cross-Platform Optimization Tools
      • Integrated Deployment Workflow
    • Zero-Cost Proxies for Rapid Filtering
    • Practical Implementation Guidelines
  • Practical Applications and Tools
    • NAS in the Real World: Integration and Deployment
    • AutoKeras: Simplicity First
    • NNI: Enterprise-Grade NAS
    • Ray Tune + Optuna: Flexible and Powerful
    • Industry Success Stories
    • From Notebook to Production: Next Steps
  • Summary
• III. AutoML for Different Data Types
• 7. AutoGluon for Tabular Data
  • Setting Up AutoGluon and Environment
    • Installation Options
    • Platform-Specific Notes
    • Setting Up Your Development Environment
    • Performance Considerations
    • Cloud Environment Recommendations
  • Choosing the Right AutoML Framework for Tabular Data
  • TabularPredictor Basics
    • Loading and Exploring Data
    • Basic Model Training
    • Understanding TabularPredictor Output
    • Different Prediction Methods
  • Binary and Multiclass Classification
    • Binary Classification in Detail
    • Multiclass Classification
  • Regression Tasks
    • Regression Versus Classification Differences
    • Interpreting Regression Performance
    • Customizing Basic Behavior
  • AutoGluons Automatic Data Processing
    • Automatic Feature Type Detection
    • Missing Value Handling
    • Categorical Encoding
  • Advanced Customization
    • Custom Hyperparameters
    • Advanced Ensemble Configuration
    • Feature Engineering Control
    • Training Process Optimization
  • Model Interpretability and Debugging
    • Interpretability Tools
    • Handling Special Data Scenarios
    • When to Use Advanced Customization
  • Project: Titanic Survival Prediction
    • Project Overview and Business Context
    • Data Exploration and Understanding
    • Baseline AutoGluon Model
    • Custom Feature Engineering for Titanic
    • Model Interpretation for Titanic
    • Performance Evaluation and Comparison
    • Model Deployment Preparation
    • Project Summary and Business Impact
    • Extending This Project
  • Data Pipeline Consistency
  • Monitoring and Maintaining Models in Production
    • Monitoring Practices
    • Monitoring Tools for AutoGluon
  • Summary
• 8. AutoML for Text and Natural Language Processing
  • AutoGluons MultiModalPredictor for Text Processing
    • Why MultiModalPredictor?
    • Underlying Model Architectures
    • Real-World Performance
  • Building Text Classification Models
    • Your First Text Classification Model
    • Understanding Model Selection
    • Hyperparameter Optimization Guidelines
  • Advanced Text Processing Capabilities
    • Beyond Classification: Advanced NLP Tasks
  • The Transformer Revolution and Beyond
  • Domain-Specific Considerations
  • Model Selection for Different Use Cases
    • Maximum Accuracy Applications
    • Balanced Applications
    • High-Throughput Applications
  • Real-World Applications and Performance
    • Industry Case Studies
    • Performance Insights
  • Production Deployment Considerations
    • Model Selection for Deployment Scenarios
    • Managed Services Versus Custom Models
    • Deploying Custom Models with SageMaker
  • Monitoring and Maintenance
    • Performance and Operational Monitoring
    • Data Drift Detection
    • Retraining and Continuous Improvement
  • Practical Project: News Article Classification
  • Summary
• 9. Time Series Forecasting with AutoGluon
  • Understanding the Time Series Challenge
  • Getting Started with TimeSeriesPredictor
  • Foundation Models and Zero-Shot Forecasting
    • The Chronos-Bolt Architecture
    • Real-World Impact of Zero-Shot Forecasting
  • Handling Complex Multiseries Scenarios
  • Advanced Capabilities: Covariate Regressors
    • Implementing Covariate Regressors
    • Business Impact of Covariate Integration
  • Model Selection and Hyperparameter Optimization
    • The Model Zoo
    • Preset Configurations
    • Custom Hyperparameter Configuration
  • Evaluation and Validation Strategies
    • Backtesting and Time-Aware Validation
    • Business-Relevant Metrics
  • Production Deployment and Cloud Integration
    • AWS Deployment Options
      • Amazon Forecast: Fully managed forecasting
      • SageMaker JumpStart: Pretrained models
      • Custom SageMaker endpoints: Maximum flexibility
      • Choosing the right deployment approach
    • Model Updating and Monitoring
      • Forecast accuracy tracking
      • Data drift detection for time series
      • SageMaker Model Monitor integration
      • Custom CloudWatch metrics and alerts
      • Automated retraining pipeline
  • Practical Project: Retail Demand Forecasting
    • Data Preparation and Exploration
    • Model Training with Advanced Features
    • Business Impact Analysis
  • Future Directions and Emerging Capabilities
  • Summary
• 10. Computer Vision with AutoGluon
  • Understanding AutoGluons Computer Vision Capabilities
    • Choosing Between Custom Models and Managed Services
    • Building Training Datasets with SageMaker Ground Truth
  • The MultiModalPredictor Advantage
  • Foundation Models Integration
    • Modern Computer Vision Architectures
    • Task Categories and Applications
  • Setting Up AutoGluon for Computer Vision
    • Installation and Environment Setup
    • Hardware Considerations
    • Verification and Basic Setup
  • Image Classification with MultiModalPredictor
    • Your First Image Classification Model
    • Understanding Data Formats and Preprocessing
    • Model Architecture Selection and Presets
    • Advanced Classification Techniques
  • Object Detection with AutoGluon
    • Understanding Object Detection
    • Basic Object Detection Setup
    • Enhanced Object Detection Capabilities
    • Advanced Object Detection Applications
  • Multimodal Computer Vision Applications
    • Combining Images with Tabular Data
    • Image and Text Integration
  • Real-World Computer Vision Project: Automated Ecommerce Product Classification
    • Project: Automated Ecommerce Product Classification
    • Data Preparation and Exploration
    • Building the Multimodal Classification System
    • Performance Analysis and Model Interpretability
    • Integration with Ecommerce Systems
  • Performance Optimization and Best Practices
    • Hardware Optimization Strategies
    • Model Monitoring and Maintenance
  • Production Deployment Considerations
    • Model Versioning and Updates
    • SageMaker Endpoint Deployment
    • SageMaker Serverless Inference for Cost-Effective Deployment
    • AWS Panorama for Edge Deployment
    • Scalable Batch Processing Service
  • Summary
• IV. Production and MLOps
• 11. Workflow Integration with MLOps Tools
  • Understanding the AutoML-MLOps Integration Landscape
    • The Scale Challenge
    • The Reproducibility Imperative
  • Experiment Tracking and Model Management
    • Hierarchical Experiment Organization
    • Artifact Management Strategies
  • Workflow Orchestration with Kubeflow
    • Designing AutoML-Aware Pipelines
    • Resource Management and Optimization
  • Production Deployment Patterns
    • Automated Validation and Quality Assurance
    • Dynamic Serving Infrastructure
    • Operational Monitoring and Maintenance
  • Monitoring and Governance
    • AutoML-Specific Monitoring Requirements
    • Governance and Compliance Frameworks
  • Integration Challenges and Solutions
    • The Artifact Explosion Challenge
    • Ensuring Reproducibility in Automated Systems
    • Bridging Technical and Business Domains
  • Best Practices and Implementation Guidelines
    • Building Progressive Capability
    • Alignment and Expectation Management
    • Risk Management and Parallel Systems
    • Observability as a Foundation
    • Organizational Learning and Adaptation
  • Summary
• 12. Data Pipeline Automation with Apache Airflow
  • Understanding Data Pipeline Requirements for AutoML
  • Airflow Architecture for Machine Learning Workflows
    • Core Components
    • Key Airflow Terminology
      • Understanding XComs
      • Understanding connections
  • Designing DAGs for AutoML Data Ingestion
  • Practical Example: Complete AutoML Data Ingestion DAG
    • DAG Initialization and Configuration
    • Understanding catchup Behavior
  • Dynamic Task Mapping for Parallel Processing
  • Feature Engineering Pipelines and Feature Stores
  • Handling Late-Arriving Data
  • Data Contracts and Schema Evolution
  • Monitoring and Data Quality Gates
  • Scaling Airflow for Enterprise AutoML
  • Operational Excellence and Best Practices
  • Summary
• 13. Deployment and Continuous Delivery for AutoML
  • The Unique Challenges of AutoML Deployment
  • Continuous Integration for Machine Learning
    • Shadow Deployment Validation
  • Continuous Deployment Pipelines
  • Testing Strategies for Automated Models
    • Contract Testing
    • Property-Based Testing
    • Metamorphic Testing
    • Adversarial Testing
  • Model Packaging and Containerization
  • Practical Example: Deploying the Adult Income Prediction Model
  • Model Serving Infrastructure
  • Monitoring and Observability in Production
    • PrometheusGrafana Monitoring Stack
    • Drift Detection with Evidently
  • Security and Compliance Considerations
    • Input Sanitization and DoS Prevention
    • Adversarial Attack Defenses
  • Continuous Learning and Feedback Loops
  • Summary
• V. Case Studies
• 14. Case Study 1: Financial ServicesReal-Time Fraud Detection at GlobalBank
  • Business Problem and Context
    • Success Criteria
  • Data Pipeline and Preparation
    • Data Pipeline Architecture
      • 1. Multisource data integration
      • 2. Data quality validation
      • 3. Time-based splits: Preventing data leakage
      • 4. Handling extreme class imbalance
    • Production Data Pipeline Considerations
  • Feature Engineering
    • 1. Temporal Features: Fraud Has a Schedule
    • 2. Velocity Features: Fraudsters Move Fast
    • 3. Behavioral Deviation: Detecting the Unusual
    • 4. Merchant Risk Scoring
    • 5. Device Trust
    • Feature Impact Summary
  • Model Development with AutoGluon
    • Sample Weighting for Cost-Sensitive Learning
    • AutoGluon Configuration
    • Why PR-AUC Instead of ROC-AUC?
    • Model Training Results
  • Model Evaluation and Interpretability
    • Finding the Optimal Threshold
    • Model Interpretability with SHAP
  • Deployment Architecture
    • FastAPI Inference Service
    • Graceful Degradation Strategy
  • Monitoring and Maintenance
    • Drift Detection with PSI
    • Automated Retraining Pipeline
    • A/B Testing for Model Updates
  • Outcomes and Lessons Learned
    • Performance Metrics
    • Business Impact
    • Key Lessons
    • The Right Division of Labor
  • Summary
• 15. Case Study 2: RetailOmnichannel Demand Forecasting
  • Business Problem and Context
    • The Scale Challenge
    • The Wake-Up Call
    • Project Objectives
  • Data Challenges: Multisource Integration
    • Point-of-Sale Data
    • Ecommerce Data
    • Inventory Data
    • Marketing and Promotions Data
    • Weather Data
    • External Signals
    • Data Pipeline Architecture
    • Key Data Decisions
  • Feature Engineering: Capturing Demand Drivers
    • Business-Aligned Metrics
    • Temporal Features (Baseline)
    • Omnichannel Behavioral Features (High Impact)
    • Weather-Driven Demand (Category-Specific)
    • Promotional Features (Complex Interactions)
    • Event-Driven Demand
    • SKU-Specific Attributes
  • Model Development: AutoGluon for Time Series at Scale
    • The AutoML Approach
    • Why Tabular AutoML for Time Series?
    • Training Strategy: Time-Based Splits
    • Multihorizon Forecasting
    • AutoGluon Configuration
    • Key Configuration Decisions
    • Handling Data Sparsity (Long-Tail SKUs)
    • Training Infrastructure
  • Evaluation: Business Metrics over Model Metrics
    • Model Performance (MAPE by Forecast Horizon)
    • Weighted Average MAPE (Business-Aligned)
    • Business Impact Metrics
      • Inventory metrics
      • Operational metrics
    • Category-Specific Performance
    • Promotional Forecast Accuracy
    • Forecast Bias Analysis
  • Deployment: Production Forecasting Pipeline
    • Pipeline Architecture
    • Technology Stack
    • Forecast Serving
    • Monitoring: Keeping Forecasts Accurate
      • Layer 1: Pipeline health (operational)
      • Layer 2: Forecast accuracy (model performance)
      • Layer 3: Business impact
    • Drift Detection
  • Business Outcomes and Lessons Learned
    • Quantified Business Impact (12 Months Postlaunch)
    • Unexpected Benefits
    • Critical Success Factors
    • What Wed Do Differently
    • Lessons for Your Demand Forecasting Project
    • When AutoML Excels for Demand Forecasting
  • Summary
• 16. Case Study 3: HealthcarePatient Readmission Prediction
  • The Business Challenge
    • The Hardest Constraint: Fairness
    • The Current State
    • Project Objectives
  • Data Challenges and HIPAA Compliance
    • Data Sources and Integration
    • Missing Data Patterns
    • Data Quality Issues
  • Feature Engineering: Structured and Unstructured Data
    • Category 1: Demographics and Social Determinants (42 Features)
    • Category 2: Clinical Complexity and Comorbidities (68 features)
    • Category 3: Utilization History (53 features)
    • Category 4: Current Encounter Features (87 features)
    • Category 5: Clinical Notes Embeddings (64 features)
      • Stage 1: Pretraining text embeddings
      • Stage 2: Combine embeddings with structured features
    • Category 6: Temporal and Interaction Features (33 features)
  • Model Development: Fairness-Aware AutoML
    • The Fairness Challenge
    • Fairness Metrics Defined
    • Baseline Model: Standard AutoGluon (No Fairness Constraints)
    • Approach 1: Remove Protected Attributes
    • Approach 2: Adversarial Debiasing
    • Approach 3 (Final): Fairness-Aware Ensemble with Reweighting
      • Technique 1: Sample reweighting
      • Technique 2: Stratified ensembling
      • Technique 3: Postprocessing calibration
    • Final Model Configuration
  • Evaluation: Performance + Fairness Metrics
    • Model Performance (Overall)
    • Fairness Metrics by Race
    • Fairness by Age Group
    • Feature Importance (Top 20 by SHAP)
    • Business Metrics
  • Deployment: Clinical Workflow Integration
    • Real-Time Prediction Architecture
    • EHR Integration (Epic)
    • Clinical Decision Support Alert
    • Care Management Workflow
    • Clinician Training and Change Management
    • Interpretability for Clinicians
    • Regulatory Considerations
  • Monitoring: Drift and Fairness
    • Three-Layer Monitoring Strategy
      • Layer 1: Model performance drift
      • Layer 2: Fairness drift
      • Layer 3: Clinical impact monitoring
    • Retraining Schedule
  • Business Outcomes and Lessons Learned
    • Clinical Outcomes
    • Fairness in Practice
    • Unexpected Benefits
    • Critical Success Factors
    • What Wed Do Differently
    • Lessons for Your Readmission Project
    • When AutoML Excels for Healthcare
  • The Production AutoML Blueprint: A Grand Synthesis
    • The Universal Patterns of Production AutoML
    • The Production Readiness Checklist
    • Final Thoughts
  • Summary
• Epilogue: The Quiet AutoML Revolution
  • The Original Promise, Delivered
  • The Assumptions We Built On
  • The Foundation Model Disruption
  • From Automation to Autonomy
  • What Agentic AutoML Actually Looks Like
  • The New Skills Gap
  • Why Everything You Learned Still Matters
  • The Governance Question Nobodys Answering
  • The Road Ahead
  • A Final Word
• Index

pokaż cały spis treści

ukryj recenzje