To improve the performance of recommender systems in dynamic user preference and time-sensitive scenarios, a matrix factorization model is proposed based on dynamic latent factors and time weighting. The model adapts to varying complexities of user and item features by dynamically adjusting the dimensionality of latent factors and incorporates a time weighting mechanism to capture user preferences more accurately. Precision@10, Recall@10, nDCG@10, and diversity are used to evaluate the model, with comparisons made against random recommendation, item-based and user-based collaborative filtering, various matrix factorization methods, and the original ALS model. Results show that the proposed model significantly outperforms baseline models in recommendation accuracy and ranking quality, with notable improvements in Precision@10 and nDCG@10. The model maintains high diversity while optimizing performance. The innovation lies in combining dynamic latent factors and time weighting mechanisms, enhancing timeliness and providing an effective approach to balance accuracy and diversity, offering new theoretical and practical insights for recommender system design and optimization.