Conversion optimization in e-commerce using neural network-based prediction of consumer behavioral patterns

Abstract

Growing volumes of data and dynamic changes in consumer behaviour require new analytical and forecasting methods to effectively estimate purchase likelihood and generate personalised recommendations. The purpose of the study is to analyse the potential of neural network methods to predict consumer behavioural patterns in e-commerce and to assess their effectiveness in optimising conversion rates. The study uses methods of systematisation and generalisation of scientific approaches, structural-functional and system analysis, comparative analysis of neural network architectures, and elements of modelling and statistical interpretation of results. Results. It is shown that neural networks provide a significant increase in conversion forecasting accuracy over traditional algorithms by accounting for complex nonlinear dependencies and contextual factors in user interactions with the platform. At the same time, the use of deep models is accompanied by increased implementation complexity, significant resource costs, and limited interpretability of results. Assessing the advantages and limitations of neural network models, in particular their ability to account for time sequences, local and global behavioural patterns, and the requirements for computing resources and data quality, enabled the determination of their feasibility for specific business scenarios. Conclusions. It is established that integrating various neural network architectures into hybrid systems increases the efficiency of forecasting users' target-action probabilities on e-commerce platforms while simultaneously accounting for structured, temporal, and graph data types and complex behavioural patterns. It is substantiated that the efficiency of user conversion forecasts depends not only on the selected model but also on the enterprise's organisational capacity: the quality of behavioural and transactional data preparation, appropriate infrastructure for model training and updating, and systematic real-time monitoring of their performance. Further research should focus on increasing the interpretability of forecasts and on developing combined methods for integrating neural networks with econometric and causal approaches.