This dissertation falls within the field of personalized and explainable recommendation systems, addressing the current limitations of artificial intelligence models, which are often perceived as “black boxes.” Although recommendation systems play a central role in many fields such as e-commerce, education, media, and entertainment, their lack of transparency, limited ability to adapt to dynamic contexts, and difficulty in leveraging heterogeneous data constitute major barriers to their adoption and user trust. The main objective of this work is to propose a unified recommendation framework combining dynamic knowledge graphs, contextual personalization, multimodal data integration, and advanced explainability. Initially, the thesis aims to design models capable of leveraging evolving knowledge graphs to capture complex semantic relationships between users, items, and contexts, while adapting to continuous changes in preferences and interactions. Techniques based on Graph Neural Networks and incremental learning will be employed to ensure efficient, real-time updating of representations. In a second phase, the work focuses on improving the explainability of recommendations through a hybrid approach combining classical explanation methods (such as SHAP or LIME) with large language models, enabling the provision of both technical justifications and natural-language explanations that are understandable to non-expert users. Finally, the thesis will address the integration of multimodal data—text, images, audio, and video—to enrich the modeling of user preferences and improve the robustness and relevance of recommendations in complex real-world environments. All of these contributions aim to lead to more transparent, adaptive, and reliable recommendation systems capable of balancing performance, interpretability, and user experience, while addressing current scientific challenges in explainable AI and knowledge graphs.