Abstract: The changes in user preference in the interaction sequence can be reflected by the changes in item attributes in the sequence. To further improve the accuracy of recommendation systems, item attributes can be integrated into sequence recommendation algorithms in a reasonable way to achieve a more precise capturing of user interests. Previous methods used multi-layer perceptron to fuse attribute information before the attention layer, which brought additional interference to the attention calculation due to the heterogeneity of item attributes and the correlation between item information and attribute information. The additional parameters brought by the multi-layer perceptron also increased the burden of the training process. This paper proposes selecting the project category that most intuitively reflects the changes in interest preferences as the attribute, separating the category interest extraction from the project prediction by decoupling the attribute representation from the project representation. The self-attention sparsification method is used to reduce the influence of noisy items in the sequence and extract more accurate category interests. Based on the extracted category interests, a multi-head cross-attention method is applied to aggregate the items in the sequence, avoiding the aggregation of items from different interest centers. Comparative and ablation experiments were conducted on real datasets such as Beauty, Sports, and Toys, and the results showed that the proposed model has better performance than baseline models and similar models.