AutoML — University Project Knowledge Base

Purpose: Deep-dive reference for learning AutoML and a structured skill set for Claude Code agents. Language: English Datasets: Iris & Wine (sklearn built-ins)

---

Learning Path

Read the files in this order for a structured understanding:

Step	File	Topic	Time
1	01_foundations.md	What is AutoML, history, motivation	~20 min
2	02_pipeline_components.md	Every stage of an AutoML pipeline	~30 min
3	03_hyperparameter_optimization.md	HPO algorithms in depth	~30 min
4	04_frameworks_comparison.md	Framework landscape & how to choose	~20 min
5	05_implementation_example.md	Hands-on code with Iris & Wine	~45 min

---

Quick-Reference Concept Map

AutoML
│
├── Problem Formulation
│   └── CASH Problem (Combined Algorithm Selection & Hyperparameter Optimization)
│
├── Pipeline Components
│   ├── Data Preprocessing       → imputation, encoding, scaling
│   ├── Feature Engineering      → selection, generation, transformation
│   ├── Algorithm Selection      → which ML model family to use
│   ├── Hyperparameter Opt.      → tuning the chosen model
│   └── Ensembling               → stacking multiple models
│
├── HPO Methods
│   ├── Grid Search              → exhaustive, expensive
│   ├── Random Search            → faster, surprisingly effective
│   ├── Bayesian Optimization    → probabilistic model of objective
│   ├── Hyperband / ASHA         → early stopping + random search
│   └── Population-based         → evolutionary strategies
│
├── Neural AutoML
│   └── NAS (Neural Architecture Search)
│       ├── Evolutionary / RL-based
│       ├── Differentiable (DARTS)
│       └── One-shot / Weight Sharing
│
└── Frameworks
    ├── Auto-sklearn   → Bayesian Opt + meta-learning, sklearn API
    ├── TPOT           → genetic programming, pipeline evolution
    ├── PyCaret        → low-code, broad coverage
    ├── H2O AutoML     → scalable, production-ready
    ├── AutoKeras      → NAS for Keras/TF models
    └── Optuna         → flexible HPO library (not full AutoML)

---

Agent Skills

For Claude Code agents, the dedicated skill file is at:

.claude/skills/automl_agent.md

This file contains structured instructions for implementing AutoML tasks programmatically.

---

Workflow

# 1 — install dependencies
uv sync

# 2 — run all experiments (saves to results/)
uv run python run_experiments.py
uv run python run_experiments.py --no-tpot   # faster, skips TPOT

# 3 — open notebooks
uv run marimo edit notebooks/01_eda.py       # EDA
uv run marimo edit notebooks/02_results.py  # Results & HPO visualizations

Optional extras:

uv sync --extra tpot        # TPOT genetic pipeline search
uv sync --extra pycaret     # PyCaret low-code AutoML
uv sync --extra viz         # Plotly + Kaleido for extra Optuna plots

Note on auto-sklearn: requires Linux/macOS + swig (sudo apt install swig). Not supported on Python 3.13 yet — use PyCaret or Optuna instead.