Regression Example

​

Dataset Overview

• MedInc: Median income in block group
• HouseAge: Median house age in block group
• AveRooms: Average number of rooms per household
• AveBedrms: Average number of bedrooms per household
• Population: Block group population
• AveOccup: Average number of household members
• Latitude: Block group latitude
• Longitude: Block group longitude

• MedHouseVal: Median house value in block group (in $100,000s)

​

Basic Usage

python -m src.main run \
  --data data/sample/california_housing.csv \
  --target MedHouseVal \
  --task regression \
  --max-iterations 3 \
  --verbose

​

What Happens

1. Data Profiling: Analyzes schema, distributions, and missing values
2. Baseline Model: Trains a simple Linear Regression to establish performance floor
3. Iteration Loop: Gemini designs, executes, and analyzes experiments
4. Report Generation: Creates a narrative Markdown report with insights

​

Expected Output

╔══════════════════════════════════════════════════════════════╗
║  ITERATION 1 - GEMINI'S REASONING                            ║
║  Thought Signature Active | Context: 4 turns                 ║
╚══════════════════════════════════════════════════════════════╝

Based on the data profile, I observe:
- Target variable (MedHouseVal) has a right-skewed distribution
- Latitude and Longitude suggest geographic patterns
- No missing values detected

For this iteration, I'm testing Random Forest with default parameters
to establish a non-linear baseline...

┌─────────────────────────────────────────────────────────────┐
│ RESULTS ANALYSIS                                            │
├─────────────────────────────────────────────────────────────┤
│ Trend: IMPROVING                                            │
│ RMSE: 0.4823   ★ NEW BEST                                   │
│   35.2% better than baseline                                │
│                                                             │
│ Key Observations:                                           │
│   - Tree-based model outperforms linear baseline            │
│   - Geographic features appear important                    │
│   - Potential for further tuning                            │
└─────────────────────────────────────────────────────────────┘

​

With Constraints

python -m src.main run \
  --data data/sample/california_housing.csv \
  --target MedHouseVal \
  --task regression \
  --constraints data/sample/constraints.md \
  --max-iterations 5 \
  --verbose

# Experiment Constraints

## Metrics
- Primary metric: RMSE

## Models
- Prefer tree-based models
- Prefer boosting methods

## Preprocessing
- Log-transform the target variable
- Use median imputation for missing values

## Termination
- Stop if no improvement for 3 iterations

​

Impact of Constraints

• Focus on tree-based models (Random Forest, XGBoost, LightGBM)
• Apply log transformation to the target variable
• Use RMSE as the primary optimization metric
• Stop early if performance plateaus

​

Advanced Configuration

python -m src.main run \
  --data data/sample/california_housing.csv \
  --target MedHouseVal \
  --task regression \
  --max-iterations 10 \
  --time-budget 7200 \
  --verbose

python -m src.main run \
  --data data/sample/california_housing.csv \
  --target MedHouseVal \
  --task regression \
  --output-dir ./my_experiment \
  --max-iterations 5

python -m src.main run \
  --resume outputs/state_20260302_143022.json

​

Interpreting Results

​

Metric Progression

• X-axis: Iteration number
• Y-axis: RMSE (lower is better)
• Trend: Ideally decreasing over iterations

​

Best Model Details

## Best Model

**Model**: XGBRegressor
**RMSE**: 0.1332
**Improvement over baseline**: 82.1%

### Hyperparameters
- learning_rate: 0.1
- max_depth: 6
- n_estimators: 100
- subsample: 0.8

### Preprocessing
- Target transformation: log
- Feature scaling: StandardScaler
- Missing value handling: median imputation

​

Key Insights Example

“Log transformation of the target variable was critical, reducing RMSE by 80%. Gradient boosting methods (XGBoost, LightGBM) consistently outperformed bagging approaches. Geographic features (Latitude, Longitude) showed high feature importance, suggesting spatial patterns in housing prices.”

​

Viewing in MLflow

mlflow ui --backend-store-uri file:./outputs/mlruns
# Open http://127.0.0.1:5000

• Compare metrics across all iterations
• View hyperparameters for each experiment
• Download saved model artifacts
• Visualize feature importance
• Export results to CSV

​

Common Results

​

Why These Results?

• Tree-based models excel: Capture non-linear relationships between features and target
• Log transformation critical: Target variable (house prices) is right-skewed
• Boosting outperforms bagging: Gradient boosting captures residual patterns
• Geographic features important: Spatial location strongly correlates with price

​

Next Steps