Jiaying Li, Jinho Yim
The proliferation of artificial intelligence (AI) and the Internet of Things (IoT) has positioned smart kitchens as a frontier for innovation in personalized nutrition, safety monitoring, and sustainable consumption. Despite rapid progress, existing approaches remain fragmented: vision-based systems struggle with occlusion, speech-driven interfaces are vulnerable to noise, and IoT sensor networks, while reliable, often lack semantic integration with user preferences. Personalized recommender systems further suffer from static designs that fail to adapt to evolving contexts. Addressing these limitations, this study introduces a multimodal deep learning framework that unifies cross-modal attention and reinforcement learning to achieve context-aware personalization. Visual, auditory, and sensor streams are embedded into a shared representation, fused via attention mechanisms, and subsequently optimized through a reinforcement learning agent that balances nutritional goals, user satisfaction, and safety requirements. Empirical evaluation across three multimodal datasets demonstrates significant improvements over strong baselines, with gains of +8.4% in Top-1 accuracy, +14.0% in F1-score for safety monitoring, and a 23.5% reduction in nutritional prediction error. Interpretability modules employing SHAP and Integrated Gradients further provide transparent explanations, enhancing trust and accountability. The findings underscore the practical value of the framework in promoting healthier diets, improving energy efficiency, and ensuring domestic safety, while laying the groundwork for future applications in healthcare, adaptive living, and sustainable human-AI interaction.