Machine Learning Techniques for Predicting User Behavior in Digital Platforms: A Systematic Review

Abstract

This systematic review examines some machine learning techniques by which prediction can be made for user behavior on a digital platform, alongside the evaluation of deep learning architectures, collaborative filtering, and other natural language processing approaches. Based on the PRISMA guidelines, we conducted a search of the literature using Google Scholar, followed by retrieving all papers from the SciSpace robust databases (n = 500) and filtering to only high-quality Q1/Q2 journal publications (n = 15). We show that session-based recommendation systems with graph neural networks (GNNs) combined with sequential models outperform conventional recurrent architectures. LSTM- and transformer-based deep learning methods exhibit superior performance over purchase behavior prediction and user engagement forecasting in many cases. The main findings provide evidence that integrating aspects of temporal dynamics, temporal positional encoding, and self-supervised learning into hybrid approaches significantly improves the prediction performance. Graph-based methods improve collaborative signals and high-order user-item interactions, whereas attention mechanisms are used to improve interpretability. However, challenges remain, including data sparsity, cold-start problems, and selection bias. Future Work: Causal Inference frameworks, Multi-Criteria Recommendation Systems, Domain-Specific Adaptations, among others. This review offers researchers and practitioners developing prediction models for digital platform use behavior extensive details across the pipeline of reconstructing predictive models of how users interact on digital platforms.