BERT has achieved tremendous success in the field of sequential recommendation. This bidirectional encoder, which models the capture of contextual information, can effectively capture the shifts in user interests. Although using BERT as an encoder is effective, we believe that there are still two limitations to using BERT in the sequential recommendation domain: User preferences for items may stem from multiple aspects, relying solely on the information within the user interaction sequence is difficult to capture the various features of the items and the user’s emerging intentions. As an encoder that models sequences information through the Mask Language Model (MLM) task, BERT struggles to model user interaction sequences that are short in length. To address these limitations, we propose the Hybrid Data Augmentation for Sequential Recommendation (HDA4SR), which uses BERT as the encoder and, in line with BERT’s characteristics, proposes two data augmentation methods from the perspectives of data-level and model-level to help the model better perform personalized user recommendations. Specifically, we propose a multi-level guided diffusion approach to introduce stable model-level data augmentation information for the target user sequence. We use other user sequences that are like the target user’s interests to introduce data-level data augmentation information for the target user sequence. Additionally, we employ contrastive learning to capture the high-order features of these two types of information, enabling the model to better model user preferences. Experiments on two datasets show that HDA4SR significantly outperforms strong baselines.