WiKAN: Lightweight Kolmogorov–Arnold Networks for accurate indoor WiFi localization

Abstract

With the growing demand for location-based services, WiFi localization plays a critical role in indoor environments. While most existing methods rely on Multi-Layer Perceptrons (MLPs), these models often suffer from limited accuracy and poor generalization across diverse deployment conditions. Kolmogorov–Arnold Networks (KANs), with their B-spline-based basis functions, better capture complex nonlinear relationships while reducing overfitting risks. However, original KANs still incur high computational costs. To address this, we propose WiKAN(WiFi KAN), a lightweight KAN-based model for indoor WiFi localization. WiKAN reduces computational complexity by simplifying the network structure to just two KANLinear layers and replacing parameter-intensive operations with optimized matrix multiplications using reconstructed basis functions. Compared to conventional computation of basis coefficients, matrix operations enable faster inference on modern hardware and improve scalability. Furthermore, WiKAN integrates SiLU and B-spline activations through a learnable linear combination, balancing smooth approximation and nonlinear representation. Experiments on three benchmark datasets (UJIIndoorLoc, Tampere, and JARIL) demonstrate that WiKAN achieves superior performance to both MLP and standard KAN models: over 99.9% building accuracy, up to 100% floor classification, and average positioning error reduced to 5.91 meters. Additionally, runtime analysis and parameter count comparisons confirm the model’s computational efficiency. Code is publicly available at: https://github.com/gyl555666/WiKAN.