A Hyperparameter-Adaptive Multilayer Perceptron Framework for Revenue Prediction Based on E-Commerce User Behavior Data

Abstract

Accurate revenue prediction remains a critical challenge for e-commerce platforms due to the highly nonlinear and dynamic nature of user behavior. At the same time, many existing machine learning approaches rely on static model configurations that limit predictive robustness. Although various techniques have been proposed for e-commerce revenue prediction, a systematic, performance-driven approach to adapting Multilayer Perceptron hyperparameters remains underexplored. This study proposes a hyperparameter-adaptive Multilayer Perceptron framework for predicting e-commerce revenue based on user behavior data. Revenue prediction is formulated as a binary classification problem, where outcomes are categorized into conversion and non-conversion events. The dataset comprises 12,330 e-commerce user sessions with behavioral and contextual features, including page interactions, session duration, bounce rate, and visitor characteristics. The proposed framework employs iterative hyperparameter adaptation by evaluating multiple MLP configurations with variations in network depth, activation functions, optimization algorithms, and regularization levels. Model performance is assessed using accuracy, precision, recall, F1-score, and Area Under the Curve. Experimental results indicate that the configuration with the Adam optimizer, ReLU activation, and moderate regularization
achieves the best performance, yielding 88.93% accuracy and an AUC of 0.91. These findings confirm that hyperparameter-adaptive selection significantly enhances prediction performance compared to static model settings. The proposed framework provides a systematic approach to improving revenue prediction accuracy and offers valuable insights for data-driven decision-making and strategic planning in e-commerce environments.