Increasing studies have shown that piRNA is closely related to the occurrence and development of a variety of complex diseases. Effective identification of disease-related piRNAs is of great significance for early diagnosis, treatment, and prevention of these diseases. However, computational methods to identify piRNA-disease association (PDA) still faces challenges such as sparse data. This paper proposed a PDA identification method based on contrastive learning, demonstrating enhanced capability in capturing latent PDA patterns. First, a multi-source feature matrix is constructed based on the representation of piRNA and disease in four feature spaces. Second, Singular Value Decomposition (SVD) view is generated to enhance node representation, and sample pairs are prepared by the original and SVD view. Finally, lightGCN model serves as encoder to aggregate neighbor information, Kolmogorov-Arnold (KAN) network performs feature mapping, and a temperature cross entropy loss function is employed to maximize the consistency of the two views to predict the PDA score. Five-fold cross-validation demonstrated that CLPDA achieved AUC and AUPR values of 98.58% and 86.25% on the piRheno dataset. Comparison experiments further confirmed the effectiveness of CLPDA has high accuracy and robustness, which helps the future application of piRNA in disease diagnosis and treatment.