Kolmogorov-Arnold Networks for Accurate Forecasting of User Actions in Complex Digital Ecosystems
Main Article Content
Abstract
In this paper, we propose a novel hybrid neural network architecture, the KAN-enhanced QResNet, specifically designed to improve user behavior prediction in complex digital ecosystems such as e-commerce platforms. The proposed model integrates the feature decomposition capabilities of Kolmogorov-Arnold Networks (KAN) with the quadratic residual interaction modeling of Quadratic Residual Networks (QResNet). This combination allows the model to capture both compositional structures and higher-order feature interactions, leading to enhanced predictive performance. We rigorously evaluated the KAN-enhanced QResNet on three diverse e-commerce datasets: E-Commerce Customer Churn Data, Multi-Category Store Behavior, and Customer Churn Analysis, demonstrating its superiority over traditional machine learning models such as Logistic Regression, Support Vector Machines (SVM), XGBoost, and standard Feedforward Neural Networks (FF-MLP). The KAN-enhanced QResNet achieved a significant F1-score of 0.86 on the E-Commerce Customer Churn Data, outperforming Logistic Regression (F1-score: 0.68) and XGBoost (F1-score: 0.77). Similarly, the model attained an F1-score of 0.79 on the Multi-Category Store Behavior dataset, compared to 0.69 achieved by XGBoost. These results demonstrate the model’s superior ability to capture complex non-linear interactions, a challenge for conventional models. Our contributions include: (1) Introducing a novel hybrid architecture that integrates KAN and QResNet to efficiently handle high-dimensional data and capture intricate feature interactions; (2) Comprehensive empirical validation across three real-world datasets, showcasing the robustness and adaptability of the model; (3) Establishing the KAN-enhanced QResNet as a state-of-the-art solution for user behavior prediction tasks, outperforming traditional models in terms of accuracy, precision, recall, and F1-score. The findings of this study highlight the transformative potential of hybrid neural network architectures for predictive modeling in complex digital environments and offer new directions for future research in domains such as customer churn prediction, personalized recommendation systems, and real-time marketing strategies.
Article Details
This work is licensed under a Creative Commons Attribution-NoDerivatives 4.0 International License.