<p>Sequential recommendation aims to predict users’ future preferences by modeling their historical interaction sequences, thereby enabling personalized and accurate recommendations. These algorithms have been widely applied in various domains, including news, advertising, and e-commerce, demonstrating significant practical value. However, real-world data often contains unreliable user behaviors. These behaviors introduce noise into the training process, distort the learning of user preferences, and ultimately degrade both accuracy and robustness. To address the robustness challenges caused by noisy interactions and data sparsity in traditional sequential recommenders, we propose IRUIRec (Robust Sequential Recommender based on Identifying and Rectifying Unreliable Instances), a robust framework leveraging a sparse attention mechanism. During training, unreliable instances are identified and rectified. At inference time, noisy interactions are effectively distinguished. Extensive experiments conducted on three real-world datasets demonstrate that the proposed model consistently outperforms state-of-the-art baselines. Furthermore, ablation studies confirm the effectiveness and necessity of each component within the proposed framework.</p>