Chapter 09: AutoML — Automated Machine Learning🔗
"AutoML democratizes ML — automating the tedious parts so experts can focus on strategy."
9.1 What is AutoML?🔗
AutoML (Automated Machine Learning) automates the process of selecting, training, and tuning machine learning models. Instead of manually trying different algorithms and hyperparameters, AutoML explores the entire search space automatically.
Manual ML vs AutoML🔗
MANUAL ML:
Data → You pick algorithm → You tune hyperparameters → You evaluate → Repeat
Time: Days to weeks
AUTOML:
Data → AutoML searches algorithms + hyperparameters → Best model returned
Time: Hours to days (with less human effort)
9.2 What AutoML Automates🔗
┌─────────────────────────────────────────────────────────┐
│ AUTOML AUTOMATION SCOPE │
│ │
│ ✅ Feature Preprocessing (encoding, scaling, imputing) │
│ ✅ Algorithm Selection (RF, XGB, SVM, Neural Net...) │
│ ✅ Hyperparameter Optimization (HPO) │
│ ✅ Model Ensembling (stack best models) │
│ ✅ Neural Architecture Search (NAS) │
│ ✅ Model Evaluation & Comparison │
│ ✅ Cross-validation strategy │
│ │
│ ❌ NOT automated: Problem definition, data collection │
│ business metrics, deployment decisions │
└─────────────────────────────────────────────────────────┘
9.3 AutoML Search Process🔗
┌──────────────────────────┐
│ AUTOML SEARCH │
│ │
Input Data ──────▶│ Algorithm Space: │
│ ├── Random Forest │
│ ├── Gradient Boosting │
│ ├── SVM │
│ ├── Neural Networks │
│ └── Linear Models │
│ │
│ Hyperparameter Space: │
│ ├── learning_rate │
│ ├── max_depth │
│ ├── n_estimators │
│ └── dropout_rate │
│ │
│ Search Strategy: │
│ ├── Bayesian Opt │
│ ├── Random Search │
│ └── Evolutionary │
└────────────┬─────────────┘
│
▼
┌───────────────┐
│ Best Model │
│ + Config │
└───────────────┘
9.4 Hyperparameter Optimization (HPO)🔗
Hyperparameters are settings you choose before training (not learned by the model).
HPO Strategies Compared🔗
GRID SEARCH: RANDOM SEARCH:
┌──┬──┬──┬──┬──┐ ┌──┬──┬──┬──┬──┐
│ │ │ │ │ │ │ │ │✓ │ │ │
├──┼──┼──┼──┼──┤ ├──┼──┼──┼──┼──┤
│ │ │ │ │ │ │ │ │ │✓ │ │
├──┼──┼──┼──┼──┤ ├──┼──┼──┼──┼──┤
│ │ │ │ │ │ │✓ │ │ │ │ │
└──┴──┴──┴──┴──┘ └──┴──┴──┴──┴──┘
Tests every combo Random samples
Slow, exhaustive Faster, misses some
BAYESIAN OPTIMIZATION (BEST):
Start → Try 5 random points → Build a model of "which areas are good"
→ Sample from good areas → Update model → Repeat
→ Smart, efficient, finds optimum faster
HPO with Optuna (Python)🔗
import optuna
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import cross_val_score
def objective(trial):
# Define hyperparameter search space
params = {
'n_estimators': trial.suggest_int('n_estimators', 50, 500),
'max_depth': trial.suggest_int('max_depth', 3, 20),
'min_samples_split': trial.suggest_int('min_samples_split', 2, 20),
'max_features': trial.suggest_categorical('max_features', ['sqrt', 'log2']),
}
clf = RandomForestClassifier(**params, random_state=42)
score = cross_val_score(clf, X_train, y_train, cv=5, scoring='accuracy').mean()
return score
# Run optimization (100 trials)
study = optuna.create_study(direction='maximize')
study.optimize(objective, n_trials=100)
print(f"Best accuracy: {study.best_value:.4f}")
print(f"Best params: {study.best_params}")
9.5 Vertex AI AutoML (GCP)🔗
Vertex AI AutoML lets you train high-quality models without writing ML code.
Supported AutoML Types🔗
| Task | What You Provide | What You Get |
|---|---|---|
| AutoML Tables | Tabular CSV data | Classification/Regression model |
| AutoML Vision | Labeled images | Image classifier |
| AutoML NLP | Labeled text | Text classifier |
| AutoML Video | Labeled videos | Video classifier |
AutoML Tables Workflow🔗
┌──────────────────────────────────────────────────────┐
│ VERTEX AI AUTOML TABLES WORKFLOW │
│ │
│ 1. Upload CSV to BigQuery or GCS │
│ │ │
│ ▼ │
│ 2. Create Dataset in Vertex AI │
│ │ │
│ ▼ │
│ 3. Configure target column + training budget │
│ (e.g., 1 hour, 8 hours, 24 hours) │
│ │ │
│ ▼ │
│ 4. AutoML trains 100s of models internally │
│ (feature engineering + HPO + ensembling) │
│ │ │
│ ▼ │
│ 5. Evaluate: AUC, Precision, Recall, Confusion Mat │
│ │ │
│ ▼ │
│ 6. Deploy best model to Vertex AI Endpoint │
│ │ │
│ ▼ │
│ 7. Online predictions via REST API │
└──────────────────────────────────────────────────────┘
Vertex AI AutoML via Python SDK🔗
from google.cloud import aiplatform
aiplatform.init(project="my-project", location="us-central1")
# Create a Dataset from BigQuery
dataset = aiplatform.TabularDataset.create(
display_name="customer-churn-dataset",
bq_source="bq://my-project.ml_datasets.customer_churn",
)
# Train AutoML Tabular model
job = aiplatform.AutoMLTabularTrainingJob(
display_name="churn-automl-v1",
optimization_prediction_type="classification", # or 'regression'
optimization_objective="maximize-au-roc",
)
model = job.run(
dataset=dataset,
target_column="churned", # column to predict
training_fraction_split=0.8,
validation_fraction_split=0.1,
test_fraction_split=0.1,
budget_milli_node_hours=1000, # 1 hour of training budget
model_display_name="churn-model-v1",
)
# Deploy
endpoint = model.deploy(machine_type="n1-standard-2")
# Predict
prediction = endpoint.predict(instances=[{
"age": 35,
"tenure_months": 12,
"monthly_charges": 65.5
}])
print(prediction.predictions)
9.6 Neural Architecture Search (NAS)🔗
NAS automates the design of neural network architectures (layer types, depths, connections).
TRADITIONAL: Data Scientist manually designs network architecture
┌────┐ ┌────┐ ┌────┐ ┌────┐ ┌────┐
│Conv│─│Conv│─│Pool│─│FC │─│Soft│ (hand-designed)
└────┘ └────┘ └────┘ └────┘ └────┘
NAS: Algorithm searches for best architecture
Search Space: [Conv3x3, Conv5x5, Pool, Skip, Dense, ...]
│
├── Try Architecture 1 → eval → 0.82 accuracy
├── Try Architecture 2 → eval → 0.85 accuracy
├── Try Architecture N → eval → 0.91 accuracy ← WINNER
└── Return best architecture
9.7 Popular AutoML Libraries🔗
| Library | Best For | Notes |
|---|---|---|
| Vertex AI AutoML | GCP production workloads | Fully managed, no code |
| H2O AutoML | Tabular data, fast | Open source |
| TPOT | Sklearn pipeline optimization | Uses genetic algorithms |
| AutoKeras | Deep learning architectures | Keras/TF-based |
| Optuna | HPO only | Flexible, widely used |
| Ray Tune | Distributed HPO | Scales to clusters |
9.8 AutoML in the MLOps Pipeline🔗
MLOps Pipeline with AutoML:
New Data Arrives
│
▼
Data Validation
│
▼
Trigger AutoML Training (Vertex AI)
┌─────────────────────────────────┐
│ AutoML internally: │
│ - Tries 100s of configs │
│ - Logs all experiments │
│ - Returns best model │
└──────────────┬──────────────────┘
│
▼
Evaluate vs current prod model
│
├── Better? → Deploy (CI/CD)
└── Worse? → Alert team
Next Chapter → 10: Experiment Tracking