A Construction Method for Personalized Movie Recommendation System Based on Embedding

Personalized recommendation systems are essential for enhancing user experience and platform profitability. This study presents a deep learning-based movie recommendation model employing embedding techniques, trained on the MovieLens-25M dataset. Unlike traditional collaborative filtering methods, the proposed system maps users and movies into a low-dimensional latent space, capturing complex nonlinear interactions through a multi-layer perceptron. To address data sparsity and cold-start limitations, negative sampling and a leave-one-out evaluation strategy were implemented. Experimental results demonstrate strong performance, achieving a Hit Ratio @10 of 0.85, thereby validating the effectiveness of embeddings in capturing user preferences. While the model shows robust recommendation quality, it is constrained by the lack of multimodal data and temporal modeling. Future work will explore Transformer architectures, large language models, and multimodal integration to enhance personalization, interpretability, and adaptability in dynamic environments.