Multi-behavior recommendation aims to utilize interactive data from multiple behaviors of users to improve recommendation performance. Existing multi-behavior recommendation methods generally directly exploit the multi-behavior data for the shared initialized user representations and involve the mining of user preferences and modeling of relationships among different behaviors in the tasks. However, these methods ignore the data imbalance under different interactive behaviors (the amount of interactive data varies greatly among different behaviors) and the information loss caused by the adaptation to the above two tasks. User preferences refer to the interests that users exhibit in different behaviors (e.g., browsing preferences), and the relationship among behaviors indicates a potential conversion from one behavior to another behavior (e.g., the conversion from browsing to purchasing). In multi-behavior recommendation, the mining of user preferences and the modeling of relationships among different behaviors can be regarded as a two-stage task. On the basis of the above considerations, the model of two-stage learning for multi-behavior recommendation (TSL-MBR for short) is proposed, which decouples the above two tasks with a two-stage strategy. In particular, the model retains the end-to-end structure and learns the two tasks by alternating training with fixed parameters. The first stage is to model user preferences under different behaviors. In this stage, the interactive data from all behaviors (without distinction as to behavior type) are first used to model the global preferences of users to alleviate the problem of data sparsity to the greatest extent. Then, the interactive data of each behavior are used to refine the behavior-specific user preference (local preference) and thus lessen the influence of the data imbalance among different behaviors. The second stage is to model the relationships among different behaviors. In this stage, the mining of user preferences and modeling of relationships among different behaviors are decoupled to relieve the information loss problem caused by adaptation to the two tasks. This two-stage model significantly improves the system’s ability to predict target behaviors. Extensive experimental results show that TSL-MBR can substantially outperform the state-of-the-art baseline models, achieving 103.01% and 33.87% of relative gains on average over the best baseline on the Tmall and Beibei datasets, respectively.