Comparison of KAN, MLP, and SVM for NIR Spectra Classification

Near infrared spectra are typically high dimensional data and widely used in various industries and scientific areas. Various machine learning algorithms have been employed to classify the NIR spectral data. This paper compares the performance of three popular models: the Kolmogorov-Arnold Network (KAN), Support Vector Machine (SVM), and Multilayer Perceptron (MLP), in classifying Mah map datasets which are derived from preprocessed NIR spectra. This study investigates the performance of three prominent machine learning models Multilayer Perceptron (MLP), Kolmogorov-Arnold Network Network (KAN), and Support Vector Machine (SVM), in classifying high-dimensional Near Infrared (NIR) data within the tobacco industry. The impact of hidden layer architecture on classification accuracy was examined for both MLP and KAN networks. While KAN demonstrated superior accuracy and stability with L-BFGS optimization, particularly when employing two hidden layers, it exhibited inferior performance with Adam optimization compared to MLP. With one hidden layer and same hidden neuron number, KAN can provide slightly better classification accuracy in high dimension than both MLP and SVM. However, SVM proved generally stable across various input dimensions, achieving high classification accuracy.