The utilization of side information in sequence recommendation alleviates to some extent the negative impact of data sparsity on recommendation performance, but acquiring and maintaining a large amount of side information is extremely costly. Furthermore, to effectively remove noise from user interaction sequences and accurately model dynamic user interests, we propose a sequence recommendation model based on hierarchical filtering and multi-contrastive learning (HCL-SRec). Firstly, filter layer is introduced to eliminate noise from user interaction sequences, enhancing the model’s focus on crucial information. Secondly, to avoid the limitation of unidirectional architectures on the hidden representation capability of behavior sequences, and the interference caused by premature integration of positional information into item representations, a decoupled bidirectional Transformer module is designed. Lastly, the interaction sequences are enhanced using a multi-pair contrastive learning to alleviate the issue of poor model performance during training caused by sparse interaction data. In which, contrastive learning combines the item representation mask and dropout mask approaches. Experimental results indicate that the algorithm outperforms others in terms of Top-K hit rate (HR@K) and normalized discounted cumulative gain (NDCG@K), further alleviating the issue of data sparsity in recommendation systems.