With the rapid development of tourism and continuous advancements in information technology, tourism recommendation systems have become crucial tools for assisting users in travel planning. Traditional collaborative filtering recommendation algorithms, while capable of providing personalized suggestions based on user historical behavior, exhibit limitations in addressing data sparsity and modeling nonlinear relationships. This paper proposes a deep learning-based collaborative filtering optimization algorithm. By introducing autoencoders and Deep Neural Networks (DNN) to perform high-dimensional nonlinear mapping of user and attraction features, combined with traditional collaborative filtering techniques, the method achieves accurate prediction of user ratings for tourist attractions. Experimental results demonstrate superior performance over conventional algorithms across metrics including precision, recall, and F1-score, offering a novel technical pathway for personalized services in smart tourism applications.