Effective negative sampling strategies can accelerate model convergence, suppress excessive randomness in negative sample generation, and enhance the predictive performance of the recommender systems modeled by implicit feedback. However, existing negative sampling methods face two potential issues: limited utilization of user-item collaborative information and treating user-negative sample interactions as driven by a singular motive, thereby ignoring the user’s multi-tiered, gradually progressive implicit preferences, which leads to low-quality negative sampling. To design negative sampling methods suited to different geometric spaces based on their data characteristics, this paper introduces a novel multi-tiered cascading negative sampling method (Multi-Tiered Cascading Negative Sampling, MTCNS) for Euclidean space graph-based collaborative filtering recommender systems. Specifically, this method processes through two cascading levels, producing high-quality overall negative sample embeddings and negative sample feature representations that embody multi-layered progressive relevance semantics. Furthermore, outside the negative sampling method, a multi-task learning framework constructs a contrastive learning auxiliary task to enhance the main task’s performance. Experiments conducted on three real datasets demonstrate that this method improves metrics such as NDCG@20 and Recall@20 by an average of over 1.5%.