Sequential recommendation predicts items in which users might be interested in based on their historical interaction data. However, in real-world recommendation scenarios, users’ historical interaction data are often sparse and inevitably contain noise generated by accidental interactions. To address these issues, this study proposes a sequential recommendation with hierarchical filtering and long- and short-term preference contrastive learning (HFCLRec). First, hierarchical filtering is employed to filter out noise in user interaction sequences. Then, bidirectional gated recurrent units and convolutional neural networks are used to capture the users’ long-term and short-term preferences, respectively. Finally, contrastive learning is introduced into HFCLRec, leveraging users’ long-term and short-term preferences to derive self-supervised signals that prevent the disruption of the inherent patterns in users’ original interaction data. Extensive experiments on the Beauty, Clothing, Sports, and ML-1M datasets validated the effectiveness of HFCLRec.