tmpnn 0.0.3

Iimplementation of TMPNN, tabular polynomial network.

TMPNN: High-Order Polynomial Regression Based on Taylor Map Factorization

This is a Tensorflow implementation of TMPNN, tabular polynomial network. TMPNN maps internal dynamical system with lower order polynomial and integrate it, resulting in high-order polynomial model with low complexity.

Example

The default usage:

from tmpnn import TMPNNRegressor, TMPNNLogisticRegressor, TMPNNClassifier, TMPNNPLTransformer

tmpnn = TMPNNRegressor(random_state=0)
tmpnn.fit(x, y)
pred = tmpnn.predict(x)
score = tmpnn.score(x, y)

However, in most cases you should scale x and optionally y:

from sklearn.compose import TransformedTargetRegressor
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import MinMaxScaler

tmpnn = TransformedTargetRegressor(
    regressor=Pipeline([
        ('mms', MinMaxScaler((-0.5,0.5))),
        ('est', TMPNNRegressor(random_state=0))
    ]),
    transformer=MinMaxScaler((-0.5,0.5))
)

TMPNN is sklearn-friendly, so you also can use it with cross validation or parameter search:

cv_scores = cross_val_score(tmpnn, x, y)

Hyperparameters

TMPNN has two major parameters:

degree: polynomial order of the Taylor Map. Default value is 2.

steps: number of iterations of applying the Taylor Map, integration steps alternatevily. Default value is 7, it might be helpful to search from 2 to 10 for each speciefic task.

In case of small datasets (n_samples < 1000) one can also tune regularizer. All tensorflow regularizers and lyapnuov (tmpnn.Lyapunov()) are acceptable. Lyapunov regularizer tend to provide models robust to extra noise.

For larger datasets one can increase number of optimization epochs max_epochs. Default value is 100.

Citation

If you use this library for a scientific publication, please use the following BibTex entry to cite our work:

@misc{ivanov2023tmpnn,
      title={TMPNN: High-Order Polynomial Regression Based on Taylor Map Factorization},
      author={Andrei Ivanov and Stefan Maria Ailuro},
      year={2023},
      eprint={2307.16105},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}