SUPR
Building Interpretable Models Using KAN
Dnr:

NAISS 2024/22-1432

Type:

NAISS Small Compute

Principal Investigator:

Amr Mehasseb Ali Alkhatib

Affiliation:

Kungliga Tekniska högskolan

Start Date:

2024-11-01

End Date:

2024-12-01

Primary Classification:

10201: Computer Sciences

Webpage:

Allocation

Abstract

Many machine learning algorithms for tabular data produce black-box models, which prevent users from understanding the rationale behind the model predictions. To overcome these limitations, an approach is proposed called ViaSHAP, which employs both the universal approximation theorem and the Kolmogorov-Arnold representation theorem to learn functions that can compute exact feature attributions using the Shapley value and also provide high predictive performance. A large-scale empirical investigation is presented, showing that the feature attributions provided by ViaSHAP align with Shapley values that are computed using a post-hoc algorithm. Furthermore, the initial results show that ViaSHAP outperforms three powerful machine learning algorithms for tabular data, Random Forests, TabNet, and XGBoost.