Information diffusion prediction in social networks aims to forecast the scale of information propagation by analyzing the paths of retweets, known as cascades. Since users often participate in multiple cascades, it is valuable to study user relationships in the entire social network, which is overlooked by some works that focus only on an individual cascade. Meanwhile, some methods underestimate the distinct contribution of each user and lack interpretability when extracting cascade features. To address these issues, this paper proposes a Multi-level Representation Learning model with a neural Hawkes process (MRLH) that captures both user relationships across cascades and spatio-temporal features within each cascade. Specifically, user representations are updated based on all the interactions in the social network, capturing global user relationships. To better extract features of each target cascade, a mutual attention mechanism is employed to distinguish user contributions in the diffusion structure, and a multi-head self-attention module with time-aware embeddings is applied to capture temporal information in the user participation sequence. Additionally, a neural Hawkes process is introduced for joint training, combining the high interpretability of the Hawkes process with the strong predictive ability of deep learning methods. Results of experiments on real-world datasets demonstrate that the proposed model outperforms state-of-the-art methods.