A Multimodal Deep Reinforcement Learning Framework for Dynamic Pricing Optimisation in E-Commerce

With the rapid expansion of e-commerce platforms, there is an increasing demand for pricing strategies that jointly optimise sales performance and profitability under highly competitive markets. To address this challenge, we propose a unified deep reinforcement learning (DRL) framework built upon ItaNet (Image, Text, and Attributes Fusion Network) for dynamic pricing in Amazon’s marketplace. By integrating multimodal product context into the DRL architecture, the framework enables pricing agents to reason over heterogeneous product characteristics and demand signals, leading to more adaptive pricing decisions. Extensive simulations across multiple DRL algorithms are conducted and compared against a range of strong baseline strategies. The results show that our approach achieves an average cumulative profit margin growth exceeding 16.7% aggregated across products and time steps, while incurring only a modest 0.3%–2% reduction in sales volume. This indicates a net improvement in total gross profit rather than gains driven by aggressive price undercutting. Overall, the findings demonstrate that ItaNet-based DRL policies deliver a more balanced and economically sustainable pricing strategy than traditional sales-driven approaches. The constructed dataset spans over 40 feature dimensions across 27 Amazon product categories, incorporating textual, visual, and numerical signals. Compared with commonly used review-centric datasets, this multimodal dataset better reflects real-world pricing environments and provides a strong foundation for future research on dynamic pricing and multimodal recommender systems.