London, United Kingdom — Nazish Ashfaq presented the conference paper titled “Balancing Performance and Complexity in KAN-based HAR Systems Using Raw and Engineered Features” at FICTA-2026, held at Holiday Inn, London, United Kingdom, during 08–09 June 2026. The paper was assigned Paper ID: 37 and was authored by Dr. Muhammad Hassan Khan, Nazish Ashfaq, and Dr. Muhammad Shahid Farid from the Department of Computer Science, University of the Punjab, Lahore, Pakistan.

The research focuses on Human Activity Recognition (HAR) using wearable sensor data and investigates how different input representations influence the performance of Kolmogorov-Arnold Network-based architectures.

The study compares two KAN-based architectures: Multi-Branch Kolmogorov-Arnold Network (MBKAN) and Multi-Branch Temporal Kolmogorov-Arnold Network (MBTKAN). The models were evaluated on two benchmark Human Activity Recognition datasets: UCI-HAR and WISDM.

The work particularly examines two data paradigms: raw sensor signals and handcrafted feature-based representations. The findings show that MBTKAN performs more effectively on raw sensor data because of its ability to capture temporal dependencies, while MBKAN achieves superior performance on handcrafted features due to its lower complexity and efficient nonlinear representation learning.

Experimental results demonstrated that MBTKAN achieved strong performance on raw sensor data, particularly on the WISDM dataset, where temporal modeling played a key role in recognizing activity patterns. In contrast, MBKAN achieved the best results on handcrafted features, obtaining 96.40% accuracy on WISDM and 95.83% accuracy on UCI-HAR.

These results highlight an important research insight: increasing architectural complexity does not always guarantee improved performance. Instead, the effectiveness of a model strongly depends on the nature of the input representation.

This presentation at FICTA-2026 emphasized the need to balance model performance, computational complexity, and data representation in practical HAR systems. The study contributes to the growing research area of interpretable and efficient deep learning models for wearable sensor-based activity recognition.

The main conclusion of the work is that MBTKAN is more suitable for raw temporal sensor signals, whereas MBKAN is more effective and computationally efficient when informative handcrafted features are available.