Most deep click-through rate (CTR) prediction models utilize a mainstream framework, which consists of the embedding layer and the feature interaction layer. Embeddings rich in semantic information directly benefit the downstream frameworks to mine potential information and achieve better performance. However, the embedding layer is rarely optimized in the CTR field. Although mapped into a low-dimensional embedding space, discrete features are still sparse. To solve this problem, we build graph structures to mine the similar interest of users and the co-occurrence relationship of items from click behavior sequences, and regard them as prior information for embedding optimization. For interpretable graph structures, we further propose graph relation embedding networks (GREENs), which utilize adapted order-wise graph convolution to alleviate the problems of data sparsity and over-smoothing. Moreover, we also propose a graph contrastive regularization module, which further normalizes graph embedding by maintaining certain graph structure information. Extensive experiments have proved that by introducing our embedding optimization methods, significant performance improvement is achieved.