Multimodal Reinforcement Learning for Dynamic Cross-Media Advertising Budget Allocation Via DDPG and PPO Combined with Meta-Learning and Adversarial Training

Abstract

This paper proposes a reinforcement learning framework to address the instability in cross-media advertising budgets, which often fail to adapt to dynamic user behavior and bidding fluctuations. The framework combines multimodal feature fusion—incorporating ad images, copy, and user behavior—and adaptive policy optimization using Deep Deterministic Policy Gradient (DDPG) and Proximal Policy Optimization (PPO). To enhance robustness across different platforms, adversarial training and meta-learning are used to adapt to shifting user feedback distributions. The policy reallocates budgets in real time, optimized through small-scale A/B testing. Experiments with over 100 million ad impressions and 5 million users show a 50% increase in ROI on social platforms, a 22.6% decrease in cost per acquisition, and near-full budget utilization on e-commerce platforms. These results highlight the effectiveness of multimodal reinforcement learning in improving cross-media resource allocation and advertising outcomes.