Neural networks have emerged as indispensable instruments for addressing intricate classification and regression challenges across numerous domains. This investigation meticulously evaluates and juxtaposes the Kolmogorov–Arnold Network (KAN), an innovative neural architecture employing B-spline activation functions derived from the principles of Kolmogorov–Arnold representation theory with the prevalent Multi-Layer Perceptron (MLP) architecture, offering a reproducible, insight-oriented baseline. Utilizing a comprehensive array of both real-world and synthetically generated datasets that exhibit diverse complexities, dimensionalities, and class distributions, this comparative study prioritizes predictive efficacy alongside model interpretability. The analytical framework integrates exhaustive preprocessing methods, multiple independent experimental trials to ensure statistical rigor, and a robust suite of evaluation metrics. Empirical evidence demonstrates that KAN models significantly surpass conventional MLPs, notably on structured datasets with relatively lower dimensionalities, by providing enhanced accuracy and interpretability despite incurring greater computational complexity and expanded parameterization. Under a standardized training protocol of 200 epochs, the KAN consistently achieves superior generalization and transparency.