Cross-domain Sequential Recommendation (CSR) is a developing task that examines how users’ sequential patterns evolve by analyzing their interactions in multiple domains. Existing CSR methods still face the following challenges: First, users’ interaction behaviors across different domains are highly complex, diverse, and challenging to predict. Second, these methods often overlook the impact of noise during the knowledge transfer process between domains. To address these challenges, we propose a novel Efficient Noise-reducing Neural Network for Cross-Domain Sequential Recommendation (ENNR). Specifically, we first employ filter encoders from a frequency domain perspective to encode the sequence data in single-domain and cross-domain, which captures the complex sequential patterns of users and improves the computational efficiency of the model. In addition, to reduce the impact of noise on model performance, we design sparse self-attention mechanisms that aim to avoid the adverse effects of the noise item at the end of the sequence, while ensuring minimal weight is assigned to randomly occurring noise items in the sequences. Through the above methods, we achieve accurate knowledge transfer, improving the overall performance of our model in CSR tasks. The final experiments on four public datasets show that the proposed ENNR surpasses state-of-the-art baselines in accuracy and efficiency.