Contrastive learning (CL) is an emerging analysis approach that aims to discover unique patterns in one dataset relative to another. By applying this approach to network analysis, we can reveal unique characteristics in one network by contrasting with another. For example, with networks of protein interactions obtained from normal and cancer tissues, we can discover unique types of interactions in cancer tissues. However, existing CL methods cannot be directly applied to networks. To address this issue, we introduce a novel approach called contrastive network representation learning (cNRL). This approach embeds network nodes into a low-dimensional space that reveals the uniqueness of one network compared to another. Within this approach, we also design a method, named i-cNRL, that offers interpretability in the learned results, allowing for understanding which specific patterns are found in one network but not the other. We demonstrate the capability of i-cNRL with multiple network models and real-world datasets. Furthermore, we provide quantitative and qualitative comparisons across i-cNRL and other potential cNRL algorithm designs.