---
description: Systematic feature engineering for ML models
triggers:
  - manual
  - ml:features
agents:
  - ml-engineer
  - data-engineer
---

# Feature Engineering Workflow

Create and manage features for ML models.

## Prerequisites
- [ ] Raw data available
- [ ] Problem understanding clear
- [ ] Feature store (optional)

## Phase 1: Feature Discovery

### Step 1.1: Exploratory Data Analysis
```yaml
agent: ml-engineer
action: analyze
analysis:
  - Data distributions
  - Missing patterns
  - Correlations
  - Target relationship
  - Temporal patterns
tools:
  - Pandas profiling
  - Sweetviz
  - Custom notebooks
```

### Step 1.2: Domain Feature Identification
```yaml
agent: ml-engineer
action: identify
sources:
  - Domain expertise
  - Literature review
  - Competitor analysis
  - Business logic
categories:
  - Identity features
  - Behavioral features
  - Temporal features
  - Aggregation features
```

## Phase 2: Feature Creation

### Step 2.1: Numerical Features
```yaml
agent: ml-engineer
action: create
transformations:
  - Scaling: StandardScaler, MinMaxScaler
  - Normalization: Log, Box-Cox, Yeo-Johnson
  - Binning: Equal-width, Equal-frequency, Custom
  - Interactions: Products, Ratios, Differences
  - Polynomial: Degree-2, Degree-3
```

### Step 2.2: Categorical Features
```yaml
agent: ml-engineer
action: create
encoding:
  - One-hot: Low cardinality
  - Target encoding: High cardinality
  - Frequency encoding: Ordinal relationships
  - Embedding: Deep learning
  - Hash encoding: Very high cardinality
```

### Step 2.3: Temporal Features
```yaml
agent: ml-engineer
action: create
features:
  - Time components: Hour, day, month, year
  - Cyclical encoding: Sin/cos transforms
  - Lag features: Previous values
  - Rolling statistics: Mean, std, min, max
  - Time since event: Days since last purchase
```

### Step 2.4: Text Features
```yaml
agent: ml-engineer
action: create
methods:
  - Bag of words: CountVectorizer
  - TF-IDF: TfidfVectorizer
  - Embeddings: Word2Vec, FastText, BERT
  - Statistics: Length, word count, sentiment
```

## Phase 3: Feature Selection

### Step 3.1: Filter Methods
```yaml
agent: ml-engineer
action: select
methods:
  - Variance threshold
  - Correlation analysis
  - Mutual information
  - Chi-square test
  - ANOVA F-test
```

### Step 3.2: Wrapper Methods
```yaml
agent: ml-engineer
action: select
methods:
  - Forward selection
  - Backward elimination
  - Recursive feature elimination
  - Exhaustive search (small sets)
```

### Step 3.3: Embedded Methods
```yaml
agent: ml-engineer
action: select
methods:
  - L1 regularization (Lasso)
  - Tree-based importance
  - Permutation importance
  - SHAP values
```

## Phase 4: Feature Validation

### Step 4.1: Quality Checks
```yaml
agent: ml-engineer
action: validate
checks:
  - No data leakage
  - Point-in-time correctness
  - Feature stability over time
  - No target leakage
  - Reproducibility
```

### Step 4.2: Performance Impact
```yaml
agent: ml-engineer
action: evaluate
experiments:
  - Baseline without feature
  - Model with feature
  - Statistical significance
  - Cross-validation stability
```

## Phase 5: Feature Documentation

### Step 5.1: Feature Catalog
```yaml
agent: ml-engineer
action: document
per_feature:
  - Name and description
  - Data type
  - Source data
  - Transformation logic
  - Business meaning
  - Owner
  - Creation date
```

### Step 5.2: Feature Store Registration
```yaml
agent: data-engineer
action: register
feature_store:
  - Feast
  - Tecton
  - Vertex AI
registration:
  - Feature definition
  - Entity mapping
  - Data source
  - Freshness requirements
  - Online/offline availability
```

## Outputs
- [ ] Feature transformations
- [ ] Feature selection results
- [ ] Feature documentation
- [ ] Feature store registration
- [ ] Performance analysis

## Quality Gates
- No data leakage verified
- Features improve model
- Point-in-time correct
- Documentation complete
- Reproducible pipeline