~ Nonlinear Adaptive Filtering as a Form of Artificial Intelligence ~

Kolmogorov-Arnold model

(for combination of quantized-continuous inputs)

Demo project code and data
The dataset BANKCHURN.CSV is taken from Neural Designer web site. It contains 10,000 records of human decisions in terms of YES/NO to the bank offers. The input is combination of quantized and continuous parameters, shown below in a table:

NameRangeType
Credit score350-850continuous
Country0-2enumeration
Gender0-1enumeration
Age18-92continuous
Tenure0-10continuous
Balance0-250898.09continuous
Product0-3enumeration
Credit card0-1enumeration
Is active0-1enumeration
Salary11.58-199992.48continuous



The functions $\phi_{q,p}$ of Kolmogorov-Arnold model were chosen according to input type either quantized or continuous.

$$ f(x_1, x_2, x_3, ... , x_n) = \sum_{q=0}^{2n} \Phi_q\left(\sum_{p=1}^{n} \phi_{q,p}(x_{p})\right), $$


The accuracy was measured by 10 fold cross-validation. About 80% of all predictions were correct. It is slightly above the accuracy reported by Neural Designer (78.9%).