Spurious link detection (SLD) aims to uncover the observed links but should not exist in a perturbed graph. Current SLD research suffers from the structure uncertainty challenge, where it is harder to identify correct spurious links that appear strong wiggly due to pollution of spurious links. In this paper, we propose a novel framework PVGCL to alleviate the widespread uncertainty reduced by spurious links with purified view modeling and contrastive learning. Firstly, we take original perturbed graph as input and Graph Convolutional Network (GCN) as polluted feature encoder to obtain the predecessor for purified view modeling. Secondly, we propose the Residual Swin Window Attention Module (RSWA) to leverage the graph invariant properties (e.g., communities and degree heterogeneity) and jointly aggregate multi-window messages of the attention mechanism. RSWA is capable of purifying already polluted feature and capturing the representative semantics of correspondingly clean graph. Thirdly, to mitigate the problem of structure uncertainty, we approximate the polluted feature and purified feature by graph contrastive learning (GCL). By doing this, the informative information of spurious links could be well preserved while the irrelevant affected features among representations could be eliminated, hence improving the discriminative capacity of the spurious link representation. The experimental results demonstrate that our approach substantially increases AUC, with relative improvements reaching as high as 12.5% over ten other competing models. These results highlight the efficacy of our method. The code is available at here .