Next-basket recommendation (NBR) seeks to suggest a set of items to users based on the sequence of their historical shopping baskets. Current NBR approaches face two major challenges: (1) Information loss: Most NBR models usually follow a two-stage framework, which first aggregates the embeddings of the items in a basket into the representation of the basket and then learns the user’s preferences based on the sequences of embedded baskets. The basket representation generation step simplifies the NBR problem, but it suffers from the loss of the detailed information about the user’s preferences for items. (2) Data sparsity: Compared to the large base of items, the interactions between a user and baskets are usually heavily sparse, making it difficult to accurately capture the user’s preferences for items. In this paper, we propose a new approach the One-Step Preference Learning model with Cross-Space Preference Transferring for NBR(OSPL-CSPT for short), which first learns the item-level and category-level preferences of users separately by a simple and effective one-step preference learning method, then transfers the category-level preferences in category space to the item space by a cross-space preference transferring module, and finally contrasts the preferences generated by transferring with the item-level preferences. The effectiveness of OSPL-CSPT has been demonstrated on real-world datasets.