Analysis of the problems of creating effective adaptive payment systems in the environment of service platforms

Abstract

Modern service platforms, particularly in car sharing and micromobility, face the challenge of customer retention and profit maximization due to their reliance on predominantly static payment models. Traditional approaches, such as fixed subscriptions or pay-per-use, fail to account for the heterogeneity of customer behavior, leading to reduced loyalty and loss of potential revenue. This study provides an in-depth analysis of the key scientific and technical problems arising in the development of adaptive payment systems capable of dynamically adjusting tariff plans to individual user behavior patterns. The core idea is to apply machine learning methods for multi-factor customer classification to create personalized payment offers that optimize both customer experience and company revenue. The paper systematically analyzes the main challenges of this process: the "cold start" problem for new users; the dynamic nature of behavioral patterns requiring the implementation of online learning mechanisms and concept drift detection; the high dimensionality of the feature space and the complexity of their engineering for accurate segmentation; ensuring model interpretability (Explainable AI) to form understandable and logical tariffs; and the architectural complexities of integrating such systems into high-load production environments. An expanded conceptual architecture for an adaptive system is proposed, consisting of modules for data collection, feature engineering, ensemble customer classification, and dynamic pricing. To solve the classification task, the application of ensemble methods such as Gradient Boosting (XGBoost, LightGBM) and Random Forest is considered in detail, as they demonstrate high accuracy on heterogeneous tabular data. A comparative analysis of the effectiveness of clustering algorithms (K-Means, DBSCAN) for initial unsupervised segmentation of the customer base is conducted. The research findings indicate that transitioning from static to adaptive pricing models based on classification can significantly enhance both Customer Satisfaction and Customer Lifetime Value. Future research will focus on developing hybrid models that combine supervised learning with Reinforcement Learning for real-time optimization of pricing strategies