Current social media content recommendation algorithms face two key problems: data sparsity and dynamic changes in user interests, which lead to poor recommendation results. This paper introduces an improved deep learning recommendation algorithm to capture the changing trend of user interests in real time and alleviate the impact of data sparsity on recommendation results. By representing users and content as nodes and interactive behaviors as edges, a user-content interaction graph is constructed. Graph Convolutional Networks (GCN) is used for feature extraction, and neighbor node information is aggregated through multi-layer convolution to generate high-order representations of users and content. A self-supervised task based on contrastive learning is designed to enhance the low-dimensional representation of users and content. The time-aware attention mechanism (TAAM) is introduced into the user behavior sequence to dynamically capture the changing trend of user interests through the time decay function and attention weight. Combining user interest prediction and content relevance, a multi-objective optimization function is designed, and a weighted sum method is used to balance the recommendation results. The TAAM-GCN algorithm in this paper performs best in the recommendation system, with an accuracy of 0.86, a diversity of 3.15, and a timeliness of 0.78, all of which are better than GNN-RL, KG-DL, ST-GCN, and CL-SR models. The experimental results verify the key role of time information in optimizing recommendation performance, provide new ideas for personalized recommendation, anomaly detection, and time series data analysis, and promote the development of social media intelligent recommendation technology.