<p>In recent years, research on multi-target cross-domain recommendation and sequential recommendation has advanced significantly, and their combination has attracted increasing attention from researchers. The challenge of cross-domain recommendation lies in how to balance the relationship between domain-common features and domain-specific features. Meanwhile, the difficulty of sequential recommendation is how to integrate item features with user representations to improve performance, which makes cross-domain sequential recommendation face greater challenges. To address the challenges in these two fields, we enhance the reusability of user representations through extracting two categories of user association representations as feature containers for cross-domain sequential recommendation. Based on this, we propose a recurrent-balance learning approach to optimize and balance the two user representations, achieving both fine-grained user modeling in sequential recommendation and balanced feature learning in cross-domain recommendation. When the number of domains increases, our method has good scalability through fine-tuning. Experimental results demonstrate the advantages of recurrent-balance learning, and our model outperforms current state-of-the-art cross-domain recommendation methods.</p>