LLM for Secure Reserve Price Optimization in Real-time Bidding

Real-time bidding (RTB) plays a crucial role in display advertising, where ad exchanges (ADXs) set reserve prices to initiate auctions among demand-side platforms (DSPs). The optimization of reserve prices is critical for publishers, as it directly impacts revenue generation and market efficiency. However, existing research predominantly assumes fixed DSP bidding strategies, which fails to account for the dynamic nature of real-world scenarios where DSP behaviors evolve due to budget constraints, market fluctuations, and competitive dynamics, etc. This discrepancy between theoretical models and practical challenges underscores the need for more effective reserve price optimization methods. In this paper, we address this gap by proposing a novel framework that integrates large language models (LLMs) into the reward shaping process of reinforcement learning (RL). All user data utilized in this study underwent rigorous anonymization and complied with GDPR-aligned privacy protocols during collection and processing. Our approach leverages the advanced comprehension and reasoning capabilities of LLMs to design and fine-tune reward structures, enabling RL algorithms to respond effectively to diverse and dynamic DSP bidding strategies. We validate our method using real-world transaction data from CAINIAO's operational environment. Security-preserving mechanisms were implemented throughout the experimental pipeline to ensure transactional data integrity and prevent unauthorized access. Experimental results demonstrate that our framework achieves a 21.51% improvement in average income compared to state-of-the-art methods.