silvermete0r/oikan

Purpose and Public Entry Points

OIKAN is a Python neuro-symbolic ML library for tabular regression and classification. Its public estimators expose scikit-learn-style fit, predict, score, formula retrieval, feature importance, and JSON persistence APIs. Package exports begin at oikan/__init__.py; estimator orchestration is centered in oikan/model.py.

Estimator Orchestration

OIKANRegressor and OIKANClassifier coordinate the end-to-end workflow: validate inputs, train/evaluate an auxiliary neural model, augment data, fit symbolic models, and expose inference/export interfaces. Classification must manage a separate symbolic model per class; regression yields one coefficient vector.

• oikan/model.py

• oikan/__init__.py

• oikan/exceptions.py

Neural Learning and Augmentation Boundary

A PyTorch tabular neural network provides learned predictions used to augment original training data with noisy/generated samples. This neural stage supports configurable hidden layers, activations, batching, epochs, learning rate, optional GPU use, and pre-symbolic evaluation. The neural model is auxiliary; exported inference is intended to rely on symbolic formulas rather than PyTorch weights.

• oikan/neural.py

• oikan/model.py

Symbolic Feature Construction and Sparse Regression

The symbolic layer expands input features into interpretable basis terms—polynomials, interactions, trigonometric, exponential, and logarithmic transforms—then applies ElasticNet regularization to select/prune terms. Coefficients and retained basis metadata form the deployable model representation.

• oikan/elasticnet.py

• oikan/utils.py

• oikan/model.py

Formula and Model Serialization

Selected basis functions plus coefficients are rendered as original symbolic strings or transformed through SymPy into simplified and LaTeX formats. Model save/load uses JSON artifacts containing feature counts, basis-function metadata, and regression coefficients or classification coefficient lists.

• oikan/model.py

• outputs/california_housing_model.json

• outputs/iris_model.json

Cross-Language Deployment Compilers

Language-specific compiler implementations consume exported symbolic-model JSON and reproduce basis evaluation plus linear prediction. Regression compilers return a scalar/vector prediction; classifier compilers compute class logits, softmax probabilities, and argmax labels. Each target language maintains distinct regression and classification implementations.

• model_compilers/python/regression_compiler.py

• model_compilers/cpp/classification_compiler.cpp

• model_compilers/rust/regression_compiler.rs

Runtime, Dependencies, and Delivery

The primary runtime is Python 3.11+ with PyTorch for neural training, NumPy for arrays, scikit-learn for ElasticNet/estimator conventions, and SymPy for formula formatting. The project is distributed as a PyPI package through pyproject.toml/setup.py; static documentation is published from docs/, with GitHub Actions static-site automation.

• pyproject.toml

• requirements.txt

• setup.py