This paper aims to design a news recommendation model based on user preferences to address the issues in recommendation systems under large datasets. Initially, four datasetsclick_history, news, news_embedding, and user_predictwere integrated into a single table, followed by data cleaning and feature engineering. Due to the large volume of data, this paper proposes necessary data filtering for the training and testing sets, utilizing temporal data to construct user feature vectors and news feature vectors. One challenge is how to effectively integrate user preferences and news features into the model to avoid overfitting or underfitting. In the model design and building phase, different methods were attempted to merge the information of users and news. Ultimately, the user preference features were processed using a fully connected layer, and the news embedding vectors were handled using an LSTM model. These two data parts were then combined into another fully connected layer, using ReLU as the activation function and CELoss as the loss function. Subsequently, the model's hyperparameters were adjusted and evaluated, achieving favorable model performance. The prediction accuracy for recommending news to users in user_predict was calculated as an evaluation criterion. Finally, this paper proposes directions for generalization and optimization in three aspects: data processing, model design, and experimental design. This includes data processing methods, potential improvements or mechanisms that could be incorporated into the model, and hyperparameter tuning. The paper primarily proposes data filtering to solve the problem of excessive data scale, which may aid in addressing recommendation system issues under large-scale datasets.