The lack of confirmed negative interactions poses a major challenge to the prediction of protein-protein interactions. The reliable selection of these negative samples within a dataset is crucial for a better understanding of the underlying patterns and dynamics. The random sampling method is the most widely used negative sampling method, where negative pairs are randomly selected from unlabelled samples (i.e., samples not experimentally confirmed as positive interactions). However, they tend to introduce inaccurately labelled negative samples, resulting in less reliable predictions, which may affect the efficiency of the learning process. Our study aims to assess the reliability of clustering-based negative sampling methods and highlight their fundamental differences from the widely used random sampling method. To achieve this goal, we propose a hierarchical clustering-based algorithm that uses different mechanisms to select negative instances from unlabelled instances. We investigated the effectiveness of our proposed approach compared to existing clustering-based negative sampling methods and random sampling on four different datasets. The results indicate that clustering-based methods surpass the commonly used random sampling method.