# Building Tabular AI: From Structured Data to Decision Intelligence **Hands-On AI Science Series — detailed book plan** **Working subtitle:** *From Structured Data to Decision Intelligence* **Target length:** 10 parts · 50 chapters · 260+ sections · labs, projects, appendices, and capstone **Audience:** advanced undergraduates, graduate students, applied ML engineers, data scientists, analytics engineers, product analysts, risk/fraud/health/finance AI practitioners **Prerequisites:** Python, pandas/SQL basics, probability, statistics, machine learning fundamentals ## Book Positioning *Building Tabular AI* treats structured data as the dominant industrial AI substrate: rows, columns, events, entities, transactions, claims, customers, products, patients, loans, experiments, and business decisions. While language, vision, and audio attract most frontier attention, most deployed AI systems still depend on tabular data. This book gives tabular AI the same scientific and engineering depth usually reserved for deep learning modalities. The book covers classical statistics, feature engineering, gradient boosting, deep tabular learning, entity embeddings, graph/tabular hybrids, AutoML, causal inference, uplift modeling, decision optimization, privacy, synthetic tabular data, explainability, monitoring, and production data systems. ## Central Thesis Tabular intelligence is the ability to convert structured records into reliable decisions under leakage, bias, shift, missingness, constraints, and human consequences. The book is organized around five verbs: 1. **Shape** — understand tables, entities, keys, time, joins, and feature construction. 2. **Predict** — build strong, calibrated models from linear baselines to boosting and deep learning. 3. **Explain** — diagnose, interpret, and audit model behavior. 4. **Decide** — turn predictions into actions, policies, experiments, and optimized interventions. 5. **Operate** — deploy, monitor, govern, and improve tabular AI in real organizations. ## What Readers Will Build Readers will build: - a reproducible tabular ML pipeline; - leakage detectors and time-aware validation; - feature engineering and feature store examples; - linear, tree, gradient boosting, and neural tabular models; - calibrated risk scoring systems; - explainability dashboards using PDP, ICE, SHAP, and counterfactuals; - causal inference and uplift modeling workflows; - fraud/risk/churn/health prediction case studies; - synthetic tabular data generators; - production monitoring for drift, quality, fairness, and calibration; - an end-to-end decision intelligence system. --- # Front Matter ## F1. Why This Book Exists Tabular data runs banks, hospitals, retailers, insurers, factories, platforms, governments, and scientific labs. Yet tabular AI is often taught as a collection of Kaggle tricks or business analytics recipes. This book presents it as a deep AI discipline with its own theory, engineering, and deployment challenges. ## F2. What This Book Covers Structured data representation, feature engineering, validation, classical and modern tabular models, missing data, imbalance, interpretability, causal decision making, synthetic data, privacy, production MLOps, and domain case studies. ## F3. Who Should Read This Book Students and practitioners who build predictive and decision systems from relational, transactional, customer, financial, clinical, operational, and experimental data. ## F4. How to Use This Book Recommended paths: - **Applied ML engineer path:** Parts I, II, III, IV, VIII, IX. - **Data science/product analytics path:** Parts I, II, III, VI, VII. - **Risk/fraud/finance path:** Parts I, III, IV, V, VII, X. - **Healthcare/regulated AI path:** Parts I, III, V, VIII, IX, X. - **Research path:** all parts plus appendices and capstone. ## F5. Running Tools and Datasets Core stack: Python, pandas, Polars, DuckDB, SQL, scikit-learn, statsmodels, XGBoost, LightGBM, CatBoost, PyTorch, pytorch-tabular, Optuna, MLflow/W&B, Great Expectations, Evidently, Feast, SHAP, DoWhy, EconML, causalml, SDV, ydata-synthetic, FastAPI, Streamlit/Gradio. Representative datasets: Adult Income, Titanic, Home Credit, Lending Club, MIMIC-derived clinical tables, UCI credit/default datasets, IEEE fraud detection, Instacart, Rossmann, airline delay, churn datasets, synthetic relational data, public A/B testing examples. --- # Part I · Foundations of Structured Data Intelligence **Part goal:** define tabular AI as structured prediction and decision making over entities, events, and relational context. ## Chapter 1. What Is Tabular AI? ### Sections 1.1 Why tables dominate deployed AI 1.2 Rows, columns, entities, events, and decisions 1.3 Prediction, ranking, scoring, segmentation, and policy 1.4 Tabular AI versus vision/language/audio AI 1.5 The full pipeline: data → features → model → decision → monitoring 1.6 Why tabular problems fail in production 1.7 Book map: shape, predict, explain, decide, operate ### Lab 1 Inspect several tabular datasets and classify task type, entity type, target definition, leakage risk, and decision use. ## Chapter 2. Tables, Schemas, and Relational Structure ### Sections 2.1 Tables as typed records 2.2 Primary keys, foreign keys, and entity resolution 2.3 Fact tables and dimension tables 2.4 One-to-one, one-to-many, and many-to-many joins 2.5 Denormalization and feature tables 2.6 Slowly changing dimensions 2.7 Data contracts and schema evolution ### Lab 2 Build a feature table from normalized relational data using SQL/DuckDB and verify key integrity. ## Chapter 3. Data Quality, Missingness, and Cleaning ### Sections 3.1 Data types and semantic types 3.2 Missing completely at random, at random, and not at random 3.3 Invalid values, outliers, duplicates 3.4 Imputation methods 3.5 Missingness as signal 3.6 Data validation tests 3.7 Cleaning without destroying signal ### Lab 3 Create a data-quality report and compare imputation strategies on downstream model performance. ## Chapter 4. Target Definition and Label Engineering ### Sections 4.1 What exactly is the model predicting? 4.2 Observation windows and prediction windows 4.3 Label delay and censoring 4.4 Proxy labels and weak supervision 4.5 Multi-class, multi-label, regression, ranking, and survival targets 4.6 Target leakage from future information 4.7 Label documentation ### Lab 4 Define churn, fraud, or risk labels from event history and show how small definition changes alter model behavior. ## Chapter 5. Validation, Splitting, and Leakage ### Sections 5.1 Random splits and when they lie 5.2 Time-based splits 5.3 Group/entity splits 5.4 Cross-validation for tabular data 5.5 Leakage through joins, aggregates, preprocessing, and duplicates 5.6 Benchmark design 5.7 Reproducible validation protocols ### Lab 5 Build leakage-safe validation for a transactional prediction task and compare it with an optimistic random split. --- # Part II · Statistical and Machine Learning Foundations **Part goal:** give the statistical spine behind tabular prediction. ## Chapter 6. Exploratory Data Analysis for Tabular AI ### Sections 6.1 Distributions and summary statistics 6.2 Correlation and association 6.3 Conditional distributions 6.4 Segment analysis 6.5 Target leakage clues 6.6 Visualizing high-cardinality categories 6.7 EDA as model debugging preparation ### Lab 6 Produce an EDA notebook that identifies data issues, predictive variables, and leakage suspects. ## Chapter 7. Linear and Generalized Linear Models ### Sections 7.1 Linear regression and least squares 7.2 Logistic regression 7.3 Regularization: ridge, lasso, elastic net 7.4 Interactions and nonlinear transformations 7.5 Generalized linear models 7.6 Interpretability and coefficient traps 7.7 Strong baselines for real work ### Lab 7 Build a regularized logistic regression risk score and interpret coefficients carefully. ## Chapter 8. Trees and Rule-Based Models ### Sections 8.1 Decision trees 8.2 Splits, impurity, and pruning 8.3 Handling categorical variables 8.4 Missing values in trees 8.5 Rule lists and rule sets 8.6 Interpretability versus instability 8.7 Trees as diagnostic models ### Lab 8 Train a decision tree and extract a compact rule-based explanation for a stakeholder. ## Chapter 9. Ensembles and Random Forests ### Sections 9.1 Bias, variance, and bagging 9.2 Random forests 9.3 Extra trees 9.4 Out-of-bag evaluation 9.5 Feature importance pitfalls 9.6 Proximities and anomaly detection 9.7 When random forests are enough ### Lab 9 Compare random forest feature importance with permutation importance under correlated features. ## Chapter 10. Gradient Boosting Machines ### Sections 10.1 Boosting intuition 10.2 Gradient boosting as functional gradient descent 10.3 XGBoost 10.4 LightGBM 10.5 CatBoost and categorical handling 10.6 Hyperparameters that matter 10.7 Why boosting is the tabular deep learning baseline ### Lab 10 Build XGBoost/LightGBM/CatBoost models and tune them with Optuna on a leakage-safe split. --- # Part III · Feature Engineering and Representation Learning **Part goal:** teach feature construction as the central craft of tabular AI. ## Chapter 11. Numeric Feature Engineering ### Sections 11.1 Scaling and transformations 11.2 Binning and monotonic transformations 11.3 Ratios, differences, and domain formulas 11.4 Aggregates and rolling statistics 11.5 Outlier treatment 11.6 Interactions 11.7 Feature documentation ### Lab 11 Engineer numeric features for a credit or churn dataset and measure incremental value. ## Chapter 12. Categorical Feature Engineering ### Sections 12.1 One-hot encoding 12.2 Ordinal encoding 12.3 Target encoding and leakage-safe variants 12.4 Hashing trick 12.5 High-cardinality categories 12.6 Rare category handling 12.7 Categorical embeddings ### Lab 12 Compare one-hot, target encoding, CatBoost handling, and embeddings for high-cardinality variables. ## Chapter 13. Time, Events, and Aggregation Features ### Sections 13.1 Event logs and temporal tables 13.2 Recency, frequency, monetary features 13.3 Rolling windows 13.4 Snapshot construction 13.5 Point-in-time correctness 13.6 Feature freshness 13.7 Temporal leakage tests ### Lab 13 Build point-in-time correct features from transactions using observation and prediction windows. ## Chapter 14. Text, Image, and External Features in Tables ### Sections 14.1 Tables with unstructured columns 14.2 Text embeddings as tabular features 14.3 Image/audio embeddings as features 14.4 Geospatial features 14.5 Knowledge-base enrichment 14.6 Feature drift from external models 14.7 Multimodal tabular systems ### Lab 14 Add text embeddings from product descriptions or notes to a tabular model and evaluate lift. ## Chapter 15. Feature Stores and Reusable Feature Pipelines ### Sections 15.1 Offline versus online features 15.2 Feature definitions and ownership 15.3 Point-in-time joins 15.4 Feature materialization 15.5 Training-serving skew 15.6 Feature monitoring 15.7 Build-versus-buy decisions ### Lab 15 Implement a minimal feature store pattern with offline training data and online-style lookup. --- # Part IV · Advanced Tabular Modeling **Part goal:** cover modern tabular models and the frontier beyond boosting. ## Chapter 16. Neural Networks for Tabular Data ### Sections 16.1 Why plain MLPs often disappoint 16.2 Normalization and regularization 16.3 Entity embeddings 16.4 Wide-and-deep models 16.5 TabNet-style attentive models 16.6 FT-Transformer and modern tabular transformers 16.7 When neural tabular models help ### Lab 16 Compare MLP, entity embedding model, and gradient boosting on mixed numerical/categorical data. ## Chapter 17. Deep Learning for Relational Tables ### Sections 17.1 Multi-table learning 17.2 Relational feature learning 17.3 Set encoders for one-to-many relations 17.4 Entity histories as sequences 17.5 Retrieval over similar entities 17.6 Relational deep learning systems 17.7 Scalability challenges ### Lab 17 Build a model that combines customer static features with a variable-length transaction history. ## Chapter 18. Graph-Enhanced Tabular AI ### Sections 18.1 When tables imply graphs 18.2 Entity graphs and transaction graphs 18.3 Graph features for tabular models 18.4 Graph neural networks 18.5 Fraud rings and relational risk 18.6 Heterogeneous graphs 18.7 Combining GNNs with boosting ### Lab 18 Construct a transaction graph and add graph-derived features to a fraud model. ## Chapter 19. AutoML and Hyperparameter Optimization ### Sections 19.1 Search spaces for tabular ML 19.2 Bayesian optimization 19.3 Multi-fidelity search 19.4 Ensembling and stacking 19.5 AutoML systems 19.6 Reproducibility and compute budgets 19.7 When AutoML is helpful or dangerous ### Lab 19 Run an AutoML or Optuna workflow and compare it with a hand-built tuned baseline. ## Chapter 20. Foundation Models for Tabular Data ### Sections 20.1 What would a tabular foundation model mean? 20.2 In-context learning over tables 20.3 Table transformers and row/column tokenization 20.4 LLMs for table reasoning 20.5 Synthetic pretraining tasks 20.6 Transfer across datasets 20.7 Current limits and research frontier ### Lab 20 Evaluate an in-context tabular model or LLM-assisted table reasoning workflow on small datasets. --- # Part V · Imbalance, Rare Events, Risk, and Anomaly Detection **Part goal:** teach the problems that dominate tabular deployment: rare positives, high cost errors, and uncertain labels. ## Chapter 21. Class Imbalance and Rare Events ### Sections 21.1 Why accuracy fails 21.2 Precision, recall, PR-AUC, ROC-AUC 21.3 Resampling and weighting 21.4 Threshold tuning 21.5 Cost-sensitive learning 21.6 Extreme imbalance 21.7 Operational alert budgets ### Lab 21 Build a rare-event classifier and choose thresholds using a cost matrix and alert budget. ## Chapter 22. Fraud and Abuse Detection ### Sections 22.1 Fraud as adaptive behavior 22.2 Transaction features 22.3 Velocity features 22.4 Graph signals 22.5 Delayed labels and chargebacks 22.6 Human review queues 22.7 Monitoring adversarial drift ### Lab 22 Build a fraud detection pipeline with delayed labels and human-review thresholding. ## Chapter 23. Credit, Risk, and Scoring Systems ### Sections 23.1 Risk scoring as tabular AI 23.2 Probability of default 23.3 Scorecards 23.4 Monotonic constraints 23.5 Calibration and expected loss 23.6 Regulatory interpretability 23.7 Reject inference and selection bias ### Lab 23 Build a calibrated credit risk score and compare logistic scorecard with gradient boosting. ## Chapter 24. Anomaly and Outlier Detection ### Sections 24.1 Outliers versus anomalies 24.2 Statistical methods 24.3 Isolation Forest and LOF 24.4 Autoencoder-based anomaly detection 24.5 Density estimation 24.6 Evaluating without ground truth 24.7 Alert explanation ### Lab 24 Build an unsupervised anomaly detector and validate it with synthetic anomalies and analyst review criteria. ## Chapter 25. Survival, Censoring, and Time-to-Event Modeling ### Sections 25.1 Time-to-event data 25.2 Censoring 25.3 Kaplan-Meier curves 25.4 Cox models 25.5 Random survival forests 25.6 Deep survival models 25.7 Applications in churn, failure, health, and credit ### Lab 25 Model time to churn or failure and compare binary prediction with survival analysis. --- # Part VI · Explainability, Debugging, and Trust **Part goal:** make tabular models understandable enough for high-stakes use. ## Chapter 26. Interpreting Tabular Models ### Sections 26.1 Global versus local explanations 26.2 Coefficients and odds ratios 26.3 Permutation importance 26.4 Partial dependence plots 26.5 ICE and ALE plots 26.6 SHAP values 26.7 Explanation failure modes ### Lab 26 Build an explanation dashboard with permutation importance, PDP, ICE, and SHAP. ## Chapter 27. Counterfactuals and Recourse ### Sections 27.1 What would need to change? 27.2 Counterfactual explanation 27.3 Actionability constraints 27.4 Recourse fairness 27.5 Optimization approaches 27.6 Human review 27.7 Counterfactual pitfalls ### Lab 27 Generate counterfactual explanations for a credit or churn model with realistic action constraints. ## Chapter 28. Model Debugging and Error Analysis ### Sections 28.1 Slice-based evaluation 28.2 Cohort analysis 28.3 Error clustering 28.4 Label audit 28.5 Feature leakage diagnosis 28.6 Data quality debugging 28.7 Debugging playbooks ### Lab 28 Create an error analysis report that identifies failure slices and proposes data/model fixes. ## Chapter 29. Fairness and Bias in Structured Data ### Sections 29.1 Sources of bias in tabular data 29.2 Group fairness metrics 29.3 Individual fairness and similarity 29.4 Proxy variables 29.5 Bias mitigation methods 29.6 Fairness under shift 29.7 Responsible reporting ### Lab 29 Evaluate a model for subgroup performance and proxy-variable effects, then test mitigation strategies. ## Chapter 30. Model Cards, Data Cards, and Governance ### Sections 30.1 Documentation as engineering 30.2 Data cards for tabular datasets 30.3 Model cards for predictive systems 30.4 Decision cards for deployed policies 30.5 Approval workflows 30.6 Audit trails 30.7 Governance for regulated domains ### Lab 30 Write a complete data card, model card, and deployment decision card for a tabular model. --- # Part VII · Causal Inference, Experiments, and Decision Intelligence **Part goal:** move beyond prediction into decisions and interventions. ## Chapter 31. Prediction Is Not Causation ### Sections 31.1 Associational versus causal questions 31.2 Confounding 31.3 Causal graphs 31.4 Interventions and counterfactuals 31.5 Selection bias 31.6 Bad controls and leakage-like causal errors 31.7 When prediction is enough ### Lab 31 Use a causal graph to explain why a strong predictive feature may be invalid for decision making. ## Chapter 32. A/B Testing and Experimentation ### Sections 32.1 Randomized experiments 32.2 Metrics and guardrails 32.3 Power and sample size 32.4 Sequential testing 32.5 Heterogeneous treatment effects 32.6 Experiment pitfalls 32.7 From experiment to policy ### Lab 32 Analyze an A/B test and compute treatment effects, confidence intervals, and guardrail metrics. ## Chapter 33. Observational Causal Inference ### Sections 33.1 Matching 33.2 Propensity scores 33.3 Inverse probability weighting 33.4 Doubly robust estimation 33.5 Difference-in-differences 33.6 Instrumental variables 33.7 Sensitivity analysis ### Lab 33 Estimate treatment effect from observational tabular data and compare methods. ## Chapter 34. Uplift Modeling and Personalized Interventions ### Sections 34.1 Treatment effect prediction 34.2 Two-model and transformed outcome approaches 34.3 Causal forests 34.4 Uplift metrics and Qini curves 34.5 Targeting policies 34.6 Budget constraints 34.7 Ethical targeting ### Lab 34 Build an uplift model for marketing or retention and select a treatment policy under budget. ## Chapter 35. Decision Optimization from Predictions ### Sections 35.1 Predictions are inputs, not decisions 35.2 Threshold policies 35.3 Utility and cost matrices 35.4 Constrained optimization 35.5 Ranking and allocation 35.6 Human-in-the-loop decision systems 35.7 Measuring decision quality ### Lab 35 Turn a risk model into an optimized review policy under capacity and fairness constraints. --- # Part VIII · Synthetic Data, Privacy, and Data-Centric AI **Part goal:** cover modern methods for data improvement, privacy protection, and synthetic structured data. ## Chapter 36. Data-Centric Tabular AI ### Sections 36.1 Why data improvements often beat model changes 36.2 Label quality 36.3 Feature quality 36.4 Data cleaning prioritization 36.5 Active learning 36.6 Weak supervision 36.7 Data-centric experiment design ### Lab 36 Improve a model by auditing labels and features rather than changing the model class. ## Chapter 37. Synthetic Tabular Data ### Sections 37.1 Why generate tabular data? 37.2 Statistical synthesizers 37.3 Copulas 37.4 GANs and VAEs for tables 37.5 Diffusion models for tabular data 37.6 Utility evaluation 37.7 Privacy and memorization risk ### Lab 37 Generate synthetic tabular data and evaluate distributional fidelity, downstream utility, and privacy risk. ## Chapter 38. Privacy-Preserving Tabular AI ### Sections 38.1 Personal and sensitive structured data 38.2 De-identification limits 38.3 Differential privacy 38.4 Federated learning for tables 38.5 Secure enclaves and clean rooms 38.6 Privacy attacks 38.7 Privacy-utility tradeoffs ### Lab 38 Train a model under privacy constraints and compare utility loss with privacy protection. ## Chapter 39. LLMs for Tabular Data Workflows ### Sections 39.1 LLMs as data assistants 39.2 Natural language to SQL 39.3 Data cleaning suggestions 39.4 Feature ideation 39.5 Automated EDA reports 39.6 LLM risks: hallucinated joins and wrong assumptions 39.7 Human-verified LLM workflows ### Lab 39 Build an LLM-assisted EDA and feature ideation workflow with verification checks. --- # Part IX · Deployment and Operations for Tabular AI **Part goal:** turn tabular models into reliable production systems. ## Chapter 40. Production Tabular ML Architecture ### Sections 40.1 Batch scoring, real-time scoring, and streaming scoring 40.2 Data warehouses, lakehouses, and feature stores 40.3 Model serving APIs 40.4 Online feature lookup 40.5 Training-serving skew 40.6 Versioning and reproducibility 40.7 Architecture patterns ### Lab 40 Deploy a tabular scoring API backed by a reproducible feature pipeline. ## Chapter 41. Monitoring Data and Model Drift ### Sections 41.1 Data drift, concept drift, label drift 41.2 Population stability index 41.3 Feature distribution monitoring 41.4 Calibration drift 41.5 Performance monitoring with delayed labels 41.6 Drift alarms and retraining triggers 41.7 Monitoring dashboards ### Lab 41 Build a drift monitoring dashboard for a deployed tabular model. ## Chapter 42. Continuous Training and MLOps ### Sections 42.1 Pipelines and orchestration 42.2 Experiment tracking 42.3 Model registry 42.4 CI/CD for ML 42.5 Backfills and historical recomputation 42.6 Safe rollout and rollback 42.7 Incident response ### Lab 42 Create an end-to-end ML pipeline with versioned data, model, metrics, and deployment artifact. ## Chapter 43. Human-in-the-Loop Tabular AI ### Sections 43.1 Review queues 43.2 Active learning loops 43.3 Analyst feedback 43.4 Decision override logging 43.5 Triage interfaces 43.6 Measuring human-AI performance 43.7 Governance and accountability ### Lab 43 Design a review queue for fraud, risk, or clinical triage with model scores and explanations. ## Chapter 44. Cost, Latency, and Reliability ### Sections 44.1 Batch cost versus online latency 44.2 Feature computation cost 44.3 Caching and precomputation 44.4 SLA design 44.5 Model complexity tradeoffs 44.6 Failure modes in production 44.7 Reliability checklists ### Lab 44 Benchmark batch and real-time scoring alternatives and choose an architecture under constraints. --- # Part X · Applications and Future Directions **Part goal:** synthesize the field across major industrial and research domains. ## Chapter 45. Finance, Credit, and Insurance ### Sections 45.1 Credit scoring 45.2 Fraud and AML 45.3 Pricing and underwriting 45.4 Claims prediction 45.5 Stress testing and macro features 45.6 Regulation and auditability 45.7 Case study architecture ### Lab 45 Design a regulated financial AI system with model, explanation, monitoring, and governance artifacts. ## Chapter 46. Healthcare and Clinical Tabular AI ### Sections 46.1 EHR tables and coding systems 46.2 Risk prediction 46.3 Clinical deterioration and readmission 46.4 Missingness and measurement bias 46.5 Patient-level validation 46.6 Clinical workflow integration 46.7 Safety and validation ### Lab 46 Build a clinical risk model with patient-level splits, calibration, and subgroup analysis. ## Chapter 47. Retail, Product, and Customer Intelligence ### Sections 47.1 Churn and retention 47.2 Lifetime value 47.3 Recommendation features 47.4 Demand and inventory signals 47.5 Pricing and promotion 47.6 Uplift and targeting 47.7 Product analytics case study ### Lab 47 Build a churn/uplift workflow that chooses customers for intervention under budget. ## Chapter 48. Operations, Manufacturing, and Supply Chain ### Sections 48.1 Operational data models 48.2 Forecasting demand with tabular features 48.3 Quality prediction 48.4 Maintenance and downtime 48.5 Optimization and scheduling 48.6 Human operations dashboards 48.7 Case study architecture ### Lab 48 Build a quality or delay prediction model and connect it to an operational decision rule. ## Chapter 49. Frontier Research in Tabular AI ### Sections 49.1 Why tabular AI remains hard 49.2 Tabular foundation models 49.3 Relational representation learning 49.4 Neural-symbolic structured data 49.5 Causal representation learning 49.6 Synthetic data and privacy frontiers 49.7 Open research problems ### Lab 49 Replicate a recent tabular model comparison and write a critical evaluation of when it beats boosting. ## Chapter 50. Capstone: End-to-End Tabular Decision Intelligence System ### Sections 50.1 Choose domain and decision 50.2 Build the dataset and target definition 50.3 Create leakage-safe features 50.4 Train baselines and strong models 50.5 Explain, calibrate, and audit 50.6 Convert predictions into decisions 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 system. - Operational delay or quality prediction system. - Synthetic tabular data generation and privacy evaluation. --- # Appendices ## Appendix A. Probability, Statistics, and Optimization Refresher Distributions, expectation, variance, likelihood, regularization, gradient descent, convexity, uncertainty, and confidence intervals. ## Appendix B. SQL, pandas, Polars, and DuckDB Primer Practical structured data manipulation, joins, windows, groupby, memory management, and reproducible pipelines. ## Appendix C. Tabular Datasets and Benchmarks Curated dataset catalog with task type, target, leakage risks, license, evaluation protocol, and baseline scores. ## Appendix D. Metrics Reference Classification, regression, ranking, calibration, imbalance, survival, uplift, fairness, business utility, and monitoring metrics. ## Appendix E. Tooling Guide scikit-learn, XGBoost, LightGBM, CatBoost, PyTorch tabular tools, Optuna, SHAP, DoWhy, EconML, SDV, Feast, Evidently, Great Expectations, MLflow. ## Appendix F. Course Syllabi Tracks: - 14-week undergraduate Tabular AI course. - 14-week graduate Decision Intelligence course. - 7-week professional tabular ML bootcamp. - 14-week regulated AI/fairness/causal inference seminar. ## Appendix G. Solutions to Selected Exercises Worked solutions for modeling, leakage diagnosis, causal inference, explainability, and deployment design. --- # Suggested 14-Week Course Track | Week | Topics | Lab | |---|---|---| | 1 | Tabular AI overview, tables, schemas | Dataset/task audit | | 2 | Data quality and missingness | Data-quality report | | 3 | Target definition and leakage | Leakage-safe labels/splits | | 4 | Linear models and trees | Baseline risk model | | 5 | Gradient boosting | XGBoost/LightGBM/CatBoost tuning | | 6 | Feature engineering | Numeric/categorical/time features | | 7 | Neural and relational tabular models | Embeddings/history model | | 8 | Imbalance, fraud, rare events | Cost-sensitive thresholding | | 9 | Explainability and debugging | SHAP/PDP/error analysis | | 10 | Fairness and governance | Model/data card | | 11 | Causal inference and A/B testing | Treatment effect analysis | | 12 | Synthetic data and privacy | Synthetic table evaluation | | 13 | Deployment and monitoring | Scoring API + drift dashboard | | 14 | Capstone presentations | Final demo | --- # Distinctive Features - Gives structured data the same depth usually reserved for neural modalities. - Treats target definition, validation, leakage, missingness, and data quality as central scientific topics. - Balances gradient boosting, deep tabular learning, causal inference, explainability, privacy, and deployment. - Connects prediction to real decisions, utility, policy, and human review. - Designed for high-stakes industrial domains where tabular AI is actually deployed.