Kolmogorov–Arnold Networks for Cross-Domain Time-Series Modeling in Health and Activity Monitoring

Hamza Haruna Mohammed, Gabriel Kiss, Frank Lindseth
Proceedings of the 7th Northern Lights Deep Learning Conference (NLDL), PMLR 307:287-306, 2026.

Abstract

Time-series data from wearable sensors and clinical assessments provide complementary perspectives on human health, yet they often remain siloed across domains. This work presents a framework for harmonizing heterogeneous time-series sources at both minute and daily resolutions, extracting interpretable temporal features through techniques such as frequency-domain analysis and automated feature engineering. On top of this feature space, we benchmark conventional machine learning methods, Random Forest, Logistic Regression, Gradient Boosting, and a Transformer baseline against a proposed Kolmogorov Arnold Networks (KANs) model, which adaptively learn functional transformations tailored to complex temporal patterns. We evaluate models on tasks including activity index prediction and disorder-related classification, with a focus on transfer learning across lifestyle and clinical domains. Results indicate that KANs achieve competitive performance and offer greater interpretability of temporal dynamics than black-box architectures. The proposed framework demonstrates how modern time-series models can enable cross-domain learning and improve the understanding of physiological and behavioral health patterns.

Cite this Paper

BibTeX
@InProceedings{pmlr-v307-mohammed26a, title = {Kolmogorov{–}Arnold Networks for Cross-Domain Time-Series Modeling in Health and Activity Monitoring}, author = {Mohammed, Hamza Haruna and Kiss, Gabriel and Lindseth, Frank}, booktitle = {Proceedings of the 7th Northern Lights Deep Learning Conference (NLDL)}, pages = {287--306}, year = {2026}, editor = {Kim, Hyeongji and Ramírez Rivera, Adín and Ricaud, Benjamin}, volume = {307}, series = {Proceedings of Machine Learning Research}, month = {06--08 Jan}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v307/main/assets/mohammed26a/mohammed26a.pdf}, url = {https://proceedings.mlr.press/v307/mohammed26a.html}, abstract = {Time-series data from wearable sensors and clinical assessments provide complementary perspectives on human health, yet they often remain siloed across domains. This work presents a framework for harmonizing heterogeneous time-series sources at both minute and daily resolutions, extracting interpretable temporal features through techniques such as frequency-domain analysis and automated feature engineering. On top of this feature space, we benchmark conventional machine learning methods, Random Forest, Logistic Regression, Gradient Boosting, and a Transformer baseline against a proposed Kolmogorov Arnold Networks (KANs) model, which adaptively learn functional transformations tailored to complex temporal patterns. We evaluate models on tasks including activity index prediction and disorder-related classification, with a focus on transfer learning across lifestyle and clinical domains. Results indicate that KANs achieve competitive performance and offer greater interpretability of temporal dynamics than black-box architectures. The proposed framework demonstrates how modern time-series models can enable cross-domain learning and improve the understanding of physiological and behavioral health patterns.} }
Endnote
%0 Conference Paper %T Kolmogorov–Arnold Networks for Cross-Domain Time-Series Modeling in Health and Activity Monitoring %A Hamza Haruna Mohammed %A Gabriel Kiss %A Frank Lindseth %B Proceedings of the 7th Northern Lights Deep Learning Conference (NLDL) %C Proceedings of Machine Learning Research %D 2026 %E Hyeongji Kim %E Adín Ramírez Rivera %E Benjamin Ricaud %F pmlr-v307-mohammed26a %I PMLR %P 287--306 %U https://proceedings.mlr.press/v307/mohammed26a.html %V 307 %X Time-series data from wearable sensors and clinical assessments provide complementary perspectives on human health, yet they often remain siloed across domains. This work presents a framework for harmonizing heterogeneous time-series sources at both minute and daily resolutions, extracting interpretable temporal features through techniques such as frequency-domain analysis and automated feature engineering. On top of this feature space, we benchmark conventional machine learning methods, Random Forest, Logistic Regression, Gradient Boosting, and a Transformer baseline against a proposed Kolmogorov Arnold Networks (KANs) model, which adaptively learn functional transformations tailored to complex temporal patterns. We evaluate models on tasks including activity index prediction and disorder-related classification, with a focus on transfer learning across lifestyle and clinical domains. Results indicate that KANs achieve competitive performance and offer greater interpretability of temporal dynamics than black-box architectures. The proposed framework demonstrates how modern time-series models can enable cross-domain learning and improve the understanding of physiological and behavioral health patterns.
APA
Mohammed, H.H., Kiss, G. & Lindseth, F.. (2026). Kolmogorov–Arnold Networks for Cross-Domain Time-Series Modeling in Health and Activity Monitoring. Proceedings of the 7th Northern Lights Deep Learning Conference (NLDL), in Proceedings of Machine Learning Research 307:287-306 Available from https://proceedings.mlr.press/v307/mohammed26a.html.

Related Material