Table of Contents

1. 1.1 Why tables dominate deployed AI
2. 1.2 Rows, columns, entities, events, and decisions
3. 1.3 Prediction, ranking, scoring, segmentation, and policy
4. 1.4 Tabular AI versus vision, language, and audio AI
5. 1.5 The full pipeline: data → features → model → decision → monitoring
6. 1.6 Why tabular problems fail in production
7. 1.7 Book map: shape, predict, explain, decide, operate
8. Lab Inspect several tabular datasets and classify task type, entity type, target definition, leakage risk, and decision use.

1. 2.1 Tables as typed records
2. 2.2 Primary keys, foreign keys, and entity resolution
3. 2.3 Fact tables and dimension tables
4. 2.4 One-to-one, one-to-many, and many-to-many joins
5. 2.5 Denormalization and feature tables
6. 2.6 Slowly changing dimensions
7. 2.7 Data contracts and schema evolution
8. Lab Build a feature table from normalized relational data using SQL/DuckDB and verify key integrity.

1. 3.1 Data types and semantic types
2. 3.2 Missing completely at random, at random, and not at random
3. 3.3 Invalid values, outliers, and duplicates
4. 3.4 Imputation methods
5. 3.5 Missingness as signal
6. 3.6 Data validation tests
7. 3.7 Cleaning without destroying signal
8. Lab Create a data-quality report and compare imputation strategies on downstream model performance.

1. 4.1 What exactly is the model predicting?
2. 4.2 Observation windows and prediction windows
3. 4.3 Label delay and censoring
4. 4.4 Proxy labels and weak supervision
5. 4.5 Multi-class, multi-label, regression, ranking, and survival targets
6. 4.6 Target leakage from future information
7. 4.7 Label documentation
8. Lab Define churn, fraud, or risk labels from event history and show how small definition changes alter model behavior.

1. 5.1 Random splits and when they lie
2. 5.2 Time-based splits
3. 5.3 Group and entity splits
4. 5.4 Cross-validation for tabular data
5. 5.5 Leakage through joins, aggregates, preprocessing, and duplicates
6. 5.6 Benchmark design
7. 5.7 Reproducible validation protocols
8. Lab Build leakage-safe validation for a transactional prediction task and compare it with an optimistic random split.

1. 6.1 Distributions and summary statistics
2. 6.2 Correlation and association
3. 6.3 Conditional distributions
4. 6.4 Segment analysis
5. 6.5 Target leakage clues
6. 6.6 Visualizing high-cardinality categories
7. 6.7 EDA as model-debugging preparation
8. Lab Produce an EDA notebook that identifies data issues, predictive variables, and leakage suspects.

1. 7.1 Linear regression and least squares
2. 7.2 Logistic regression
3. 7.3 Regularization: ridge, lasso, and elastic net
4. 7.4 Interactions and nonlinear transformations
5. 7.5 Generalized linear models
6. 7.6 Interpretability and coefficient traps
7. 7.7 Strong baselines for real work
8. Lab Build a regularized logistic regression risk score and interpret coefficients carefully.

1. 8.1 Decision trees
2. 8.2 Splits, impurity, and pruning
3. 8.3 Handling categorical variables
4. 8.4 Missing values in trees
5. 8.5 Rule lists and rule sets
6. 8.6 Interpretability versus instability
7. 8.7 Trees as diagnostic models
8. Lab Train a decision tree and extract a compact rule-based explanation for a stakeholder.

1. 9.1 Bias, variance, and bagging
2. 9.2 Random forests
3. 9.3 Extra trees
4. 9.4 Out-of-bag evaluation
5. 9.5 Feature-importance pitfalls
6. 9.6 Proximities and anomaly detection
7. 9.7 When random forests are enough
8. Lab Compare random forest feature importance with permutation importance under correlated features.

1. 10.1 Boosting intuition
2. 10.2 Gradient boosting as functional gradient descent
3. 10.3 XGBoost
4. 10.4 LightGBM
5. 10.5 CatBoost and categorical handling
6. 10.6 Hyperparameters that matter
7. 10.7 Why boosting is the tabular deep-learning baseline
8. Lab Build XGBoost, LightGBM, and CatBoost models and tune them with Optuna on a leakage-safe split.

1. 11.1 Scaling and transformations
2. 11.2 Binning and monotonic transformations
3. 11.3 Ratios, differences, and domain formulas
4. 11.4 Aggregates and rolling statistics
5. 11.5 Outlier treatment
6. 11.6 Interactions
7. 11.7 Feature documentation
8. Lab Engineer numeric features for a credit or churn dataset and measure incremental value.

1. 12.1 One-hot encoding
2. 12.2 Ordinal encoding
3. 12.3 Target encoding and leakage-safe variants
4. 12.4 The hashing trick
5. 12.5 High-cardinality categories
6. 12.6 Rare category handling
7. 12.7 Categorical embeddings
8. Lab Compare one-hot, target encoding, CatBoost handling, and embeddings for high-cardinality variables.

1. 13.1 Event logs and temporal tables
2. 13.2 Recency, frequency, and monetary features
3. 13.3 Rolling windows
4. 13.4 Snapshot construction
5. 13.5 Point-in-time correctness
6. 13.6 Feature freshness
7. 13.7 Temporal leakage tests
8. Lab Build point-in-time correct features from transactions using observation and prediction windows.

1. 14.1 Tables with unstructured columns
2. 14.2 Text embeddings as tabular features
3. 14.3 Image and audio embeddings as features
4. 14.4 Geospatial features
5. 14.5 Knowledge-base enrichment
6. 14.6 Feature drift from external models
7. 14.7 Multimodal tabular systems
8. Lab Add text embeddings from product descriptions or notes to a tabular model and evaluate lift.

1. 15.1 Offline versus online features
2. 15.2 Feature definitions and ownership
3. 15.3 Point-in-time joins
4. 15.4 Feature materialization
5. 15.5 Training-serving skew
6. 15.6 Feature monitoring
7. 15.7 Build-versus-buy decisions
8. Lab Implement a minimal feature store pattern with offline training data and online-style lookup.

1. 16.1 Why plain MLPs often disappoint
2. 16.2 Normalization and regularization
3. 16.3 Entity embeddings
4. 16.4 Wide-and-deep models
5. 16.5 TabNet-style attentive models
6. 16.6 FT-Transformer and modern tabular transformers
7. 16.7 When neural tabular models help
8. Lab Compare an MLP, an entity-embedding model, and gradient boosting on mixed numerical/categorical data.

1. 17.1 Multi-table learning
2. 17.2 Relational feature learning
3. 17.3 Set encoders for one-to-many relations
4. 17.4 Entity histories as sequences
5. 17.5 Retrieval over similar entities
6. 17.6 Relational deep-learning systems
7. 17.7 Scalability challenges
8. Lab Build a model that combines customer static features with a variable-length transaction history.

1. 18.1 When tables imply graphs
2. 18.2 Entity graphs and transaction graphs
3. 18.3 Graph features for tabular models
4. 18.4 Graph neural networks
5. 18.5 Fraud rings and relational risk
6. 18.6 Heterogeneous graphs
7. 18.7 Combining GNNs with boosting
8. Lab Construct a transaction graph and add graph-derived features to a fraud model.

1. 19.1 Search spaces for tabular ML
2. 19.2 Bayesian optimization
3. 19.3 Multi-fidelity search
4. 19.4 Ensembling and stacking
5. 19.5 AutoML systems
6. 19.6 Reproducibility and compute budgets
7. 19.7 When AutoML is helpful or dangerous
8. Lab Run an AutoML or Optuna workflow and compare it with a hand-built tuned baseline.

1. 20.1 What would a tabular foundation model mean?
2. 20.2 In-context learning over tables
3. 20.3 Table transformers and row/column tokenization
4. 20.4 LLMs for table reasoning
5. 20.5 Synthetic pretraining tasks
6. 20.6 Transfer across datasets
7. 20.7 Current limits and the research frontier
8. Lab Evaluate an in-context tabular model or LLM-assisted table-reasoning workflow on small datasets.

1. 21.1 Why accuracy fails
2. 21.2 Precision, recall, PR-AUC, and ROC-AUC
3. 21.3 Resampling and weighting
4. 21.4 Threshold tuning
5. 21.5 Cost-sensitive learning
6. 21.6 Extreme imbalance
7. 21.7 Operational alert budgets
8. Lab Build a rare-event classifier and choose thresholds using a cost matrix and alert budget.

1. 22.1 Fraud as adaptive behavior
2. 22.2 Transaction features
3. 22.3 Velocity features
4. 22.4 Graph signals
5. 22.5 Delayed labels and chargebacks
6. 22.6 Human review queues
7. 22.7 Monitoring adversarial drift
8. Lab Build a fraud detection pipeline with delayed labels and human-review thresholding.

1. 23.1 Risk scoring as tabular AI
2. 23.2 Probability of default
3. 23.3 Scorecards
4. 23.4 Monotonic constraints
5. 23.5 Calibration and expected loss
6. 23.6 Regulatory interpretability
7. 23.7 Reject inference and selection bias
8. Lab Build a calibrated credit risk score and compare a logistic scorecard with gradient boosting.

1. 24.1 Outliers versus anomalies
2. 24.2 Statistical methods
3. 24.3 Isolation Forest and LOF
4. 24.4 Autoencoder-based anomaly detection
5. 24.5 Density estimation
6. 24.6 Evaluating without ground truth
7. 24.7 Alert explanation
8. Lab Build an unsupervised anomaly detector and validate it with synthetic anomalies and analyst review criteria.

1. 25.1 Time-to-event data
2. 25.2 Censoring
3. 25.3 Kaplan-Meier curves
4. 25.4 Cox models
5. 25.5 Random survival forests
6. 25.6 Deep survival models
7. 25.7 Applications in churn, failure, health, and credit
8. Lab Model time to churn or failure and compare binary prediction with survival analysis.

1. 26.1 Global versus local explanations
2. 26.2 Coefficients and odds ratios
3. 26.3 Permutation importance
4. 26.4 Partial dependence plots
5. 26.5 ICE and ALE plots
6. 26.6 SHAP values
7. 26.7 Explanation failure modes
8. Lab Build an explanation dashboard with permutation importance, PDP, ICE, and SHAP.

1. 27.1 What would need to change?
2. 27.2 Counterfactual explanation
3. 27.3 Actionability constraints
4. 27.4 Recourse fairness
5. 27.5 Optimization approaches
6. 27.6 Human review
7. 27.7 Counterfactual pitfalls
8. Lab Generate counterfactual explanations for a credit or churn model with realistic action constraints.

1. 28.1 Slice-based evaluation
2. 28.2 Cohort analysis
3. 28.3 Error clustering
4. 28.4 Label audit
5. 28.5 Feature leakage diagnosis
6. 28.6 Data quality debugging
7. 28.7 Debugging playbooks
8. Lab Create an error-analysis report that identifies failure slices and proposes data and model fixes.

1. 29.1 Sources of bias in tabular data
2. 29.2 Group fairness metrics
3. 29.3 Individual fairness and similarity
4. 29.4 Proxy variables
5. 29.5 Bias mitigation methods
6. 29.6 Fairness under shift
7. 29.7 Responsible reporting
8. Lab Evaluate a model for subgroup performance and proxy-variable effects, then test mitigation strategies.

1. 30.1 Documentation as engineering
2. 30.2 Data cards for tabular datasets
3. 30.3 Model cards for predictive systems
4. 30.4 Decision cards for deployed policies
5. 30.5 Approval workflows
6. 30.6 Audit trails
7. 30.7 Governance for regulated domains
8. Lab Write a complete data card, model card, and deployment decision card for a tabular model.

1. 31.1 Associational versus causal questions
2. 31.2 Confounding
3. 31.3 Causal graphs
4. 31.4 Interventions and counterfactuals
5. 31.5 Selection bias
6. 31.6 Bad controls and leakage-like causal errors
7. 31.7 When prediction is enough
8. Lab Use a causal graph to explain why a strong predictive feature may be invalid for decision-making.

1. 32.1 Randomized experiments
2. 32.2 Metrics and guardrails
3. 32.3 Power and sample size
4. 32.4 Sequential testing
5. 32.5 Heterogeneous treatment effects
6. 32.6 Experiment pitfalls
7. 32.7 From experiment to policy
8. Lab Analyze an A/B test and compute treatment effects, confidence intervals, and guardrail metrics.

1. 33.1 Matching
2. 33.2 Propensity scores
3. 33.3 Inverse probability weighting
4. 33.4 Doubly robust estimation
5. 33.5 Difference-in-differences
6. 33.6 Instrumental variables
7. 33.7 Sensitivity analysis
8. Lab Estimate a treatment effect from observational tabular data and compare methods.

1. 34.1 Treatment effect prediction
2. 34.2 Two-model and transformed-outcome approaches
3. 34.3 Causal forests
4. 34.4 Uplift metrics and Qini curves
5. 34.5 Targeting policies
6. 34.6 Budget constraints
7. 34.7 Ethical targeting
8. Lab Build an uplift model for marketing or retention and select a treatment policy under budget.

1. 35.1 Predictions are inputs, not decisions
2. 35.2 Threshold policies
3. 35.3 Utility and cost matrices
4. 35.4 Constrained optimization
5. 35.5 Ranking and allocation
6. 35.6 Human-in-the-loop decision systems
7. 35.7 Measuring decision quality
8. Lab Turn a risk model into an optimized review policy under capacity and fairness constraints.

1. 36.1 Why data improvements often beat model changes
2. 36.2 Label quality
3. 36.3 Feature quality
4. 36.4 Data cleaning prioritization
5. 36.5 Active learning
6. 36.6 Weak supervision
7. 36.7 Data-centric experiment design
8. Lab Improve a model by auditing labels and features rather than changing the model class.

1. 37.1 Why generate tabular data?
2. 37.2 Statistical synthesizers
3. 37.3 Copulas
4. 37.4 GANs and VAEs for tables
5. 37.5 Diffusion models for tabular data
6. 37.6 Utility evaluation
7. 37.7 Privacy and memorization risk
8. Lab Generate synthetic tabular data and evaluate distributional fidelity, downstream utility, and privacy risk.

1. 38.1 Personal and sensitive structured data
2. 38.2 De-identification limits
3. 38.3 Differential privacy
4. 38.4 Federated learning for tables
5. 38.5 Secure enclaves and clean rooms
6. 38.6 Privacy attacks
7. 38.7 Privacy-utility tradeoffs
8. Lab Train a model under privacy constraints and compare utility loss with privacy protection.

1. 39.1 LLMs as data assistants
2. 39.2 Natural language to SQL
3. 39.3 Data cleaning suggestions
4. 39.4 Feature ideation
5. 39.5 Automated EDA reports
6. 39.6 LLM risks: hallucinated joins and wrong assumptions
7. 39.7 Human-verified LLM workflows
8. Lab Build an LLM-assisted EDA and feature-ideation workflow with verification checks.

1. 40.1 Batch scoring, real-time scoring, and streaming scoring
2. 40.2 Data warehouses, lakehouses, and feature stores
3. 40.3 Model serving APIs
4. 40.4 Online feature lookup
5. 40.5 Training-serving skew
6. 40.6 Versioning and reproducibility
7. 40.7 Architecture patterns
8. Lab Deploy a tabular scoring API backed by a reproducible feature pipeline.

1. 41.1 Data drift, concept drift, and label drift
2. 41.2 Population stability index
3. 41.3 Feature distribution monitoring
4. 41.4 Calibration drift
5. 41.5 Performance monitoring with delayed labels
6. 41.6 Drift alarms and retraining triggers
7. 41.7 Monitoring dashboards
8. Lab Build a drift monitoring dashboard for a deployed tabular model.

1. 42.1 Pipelines and orchestration
2. 42.2 Experiment tracking
3. 42.3 Model registry
4. 42.4 CI/CD for ML
5. 42.5 Backfills and historical recomputation
6. 42.6 Safe rollout and rollback
7. 42.7 Incident response
8. Lab Create an end-to-end ML pipeline with versioned data, model, metrics, and deployment artifact.

1. 43.1 Review queues
2. 43.2 Active learning loops
3. 43.3 Analyst feedback
4. 43.4 Decision override logging
5. 43.5 Triage interfaces
6. 43.6 Measuring human-AI performance
7. 43.7 Governance and accountability
8. Lab Design a review queue for fraud, risk, or clinical triage with model scores and explanations.

1. 44.1 Batch cost versus online latency
2. 44.2 Feature computation cost
3. 44.3 Caching and precomputation
4. 44.4 SLA design
5. 44.5 Model complexity tradeoffs
6. 44.6 Failure modes in production
7. 44.7 Reliability checklists
8. Lab Benchmark batch and real-time scoring alternatives and choose an architecture under constraints.

1. 45.1 Credit scoring
2. 45.2 Fraud and AML
3. 45.3 Pricing and underwriting
4. 45.4 Claims prediction
5. 45.5 Stress testing and macro features
6. 45.6 Regulation and auditability
7. 45.7 Case study architecture
8. Lab Design a regulated financial AI system with model, explanation, monitoring, and governance artifacts.

1. 46.1 EHR tables and coding systems
2. 46.2 Risk prediction
3. 46.3 Clinical deterioration and readmission
4. 46.4 Missingness and measurement bias
5. 46.5 Patient-level validation
6. 46.6 Clinical workflow integration
7. 46.7 Safety and validation
8. Lab Build a clinical risk model with patient-level splits, calibration, and subgroup analysis.

1. 47.1 Churn and retention
2. 47.2 Lifetime value
3. 47.3 Recommendation features
4. 47.4 Demand and inventory signals
5. 47.5 Pricing and promotion
6. 47.6 Uplift and targeting
7. 47.7 Product analytics case study
8. Lab Build a churn/uplift workflow that chooses customers for intervention under budget.

1. 48.1 Operational data models
2. 48.2 Forecasting demand with tabular features
3. 48.3 Quality prediction
4. 48.4 Maintenance and downtime
5. 48.5 Optimization and scheduling
6. 48.6 Human operations dashboards
7. 48.7 Case study architecture
8. Lab Build a quality or delay prediction model and connect it to an operational decision rule.

1. 49.1 Why tabular AI remains hard
2. 49.2 Tabular foundation models
3. 49.3 Relational representation learning
4. 49.4 Neural-symbolic structured data
5. 49.5 Causal representation learning
6. 49.6 Synthetic data and privacy frontiers
7. 49.7 Open research problems
8. Lab Replicate a recent tabular model comparison and write a critical evaluation of when it beats boosting.

1. 50.1 Choose the domain and decision
2. 50.2 Build the dataset and target definition
3. 50.3 Create leakage-safe features
4. 50.4 Train baselines and strong models
5. 50.5 Explain, calibrate, and audit
6. 50.6 Convert predictions into decisions
7. 50.7 Deploy, monitor, and document

Capstone options: credit/risk scoring system · fraud detection and review queue · churn prediction plus uplift targeting · clinical risk prediction · operational delay or quality prediction · synthetic tabular data generation and privacy evaluation.