pyoccam 0.9.4

OCCAM Reconstructability Analysis Tools

PyOccam

Python bindings for OCCAM Reconstructability Analysis

What is PyOccam?

PyOccam wraps the OCCAM Reconstructability Analysis tools in a modern Python package. It provides information-theoretic modeling for categorical data, useful for:

• Environmental modeling (landslides, wildfires)
• Biomedical research (disease risk factors)
• Social science and survey analysis
• Any domain with discrete/categorical variables

OCCAM discovers significant variable relationships while emphasizing interpretability over black-box prediction.

Installation

pip install pyoccam

Quick Start

import pyoccam

# Load built-in dataset
data = pyoccam.load_dementia()
print(f"Loaded {data.n_samples} samples, {data.n_features} features")

# Run search to find best model
manager = data.manager
report = manager.generate_search_report("loopless-up", levels=5, width=3)
print(report)

# Get best model by BIC
best = manager.get_best_model_by_bic()
print(f"Best model: {best}")

# Generate fit report with confusion matrix
fit_report = manager.generate_fit_report(best, target_state="0")
print(fit_report)

# Get confusion matrix as dictionary
cm = manager.get_confusion_matrix(best, target_state="0")
print(f"Accuracy: {cm['train_accuracy']:.1%}")

Convert Your Own CSV Data

import pyoccam

# Convert CSV to OCCAM format with train/test split
output_file, data = pyoccam.make_occam_input_from_csv(
    "mydata.csv",
    test_split=0.2,                  # 20% held out for validation
    random_state=42,                 # Reproducible split
    max_cardinality=20,              # Exclude columns with >20 unique values
    dv_column="target",              # Specify dependent variable
    exclude_columns=["ID", "Name"]   # Always exclude these
)

# Analyze - now with train AND test metrics!
best = data.quick_search()
cm = data.manager.get_confusion_matrix(best, target_state="0")
print(f"Train accuracy: {cm['train_accuracy']:.1%}")
print(f"Test accuracy:  {cm['test_accuracy']:.1%}")

Or from the command line:

# With 20% test split
python -m pyoccam csv2occam mydata.csv --test-split 0.2 --exclude ID,Name

Model Selection Methods

# Different criteria for "best" model:
best_bic = manager.get_best_model_by_bic()           # Most parsimonious (recommended)
best_aic = manager.get_best_model_by_aic()           # Less penalty for complexity  
best_info = manager.get_best_model_by_information()  # Highest info, alpha < 0.05

Built-in Datasets

data = pyoccam.load_dementia()    # Alzheimer's disease risk factors
data = pyoccam.load_landslides()  # Geological hazard data
data = pyoccam.load_data("file.txt")  # Any OCCAM format file

Search Types

• "loopless-up" - Bottom-up search, no feedback loops (recommended for directed systems)
• "loopless-down" - Top-down search, no loops
• "full-up" - Bottom-up with loops allowed
• "full-down" - Top-down with loops allowed

Links

• Practical Guide: PRACTICAL_GUIDE.md - Tips from real projects
• OCCAM Manual: PDF - Complete theory & reference
• Source Code: github.com/occam-ra/occam
• Web Interface: occam.hsd.pdx.edu

License

GPL v3 - See LICENSE file

Credits

OCCAM was developed at Portland State University by Prof. Martin Zwick and contributors including Ken Willett, Joe Fusion, and H. Forrest Alexander. Python bindings by David Percy.