Graph Neural Networks (GNNs) have emerged as a prominent approach in Recommendation Systems (RSs) owing to their effectiveness in modeling high-order interactions among nodes. Knowledge Graphs (KGs), which contain rich semantic information about entities and their relationships, can enhance the accuracy and transparency of recommendations when integrated with user-item relationship graphs. Despite the contributions of KGs to the advancement of RSs, several limitations remain. First, the scarcity of labeled data can hinder the ability of GNNs to effectively learn the representation vectors of users and items. Second, traditional self-supervised learning model structures tend to be overly complex, resulting in poor generalization and scalability. To tackle the aforementioned challenges, a simplified and enhanced Multi-Graph Contrastive Learning Network, referred to as LightMGCL, is introduced. Inspired by self-supervised learning, we employ an edge dropout technique to alter the graph structure, thereby generating multiple augmented views of both the KG and the user-item relationship graph. These augmented views are then jointly trained alongside the original supervised task. Through theoretical analysis, we demonstrate that LightMGCL can effectively address data sparsity and label missingness in RSs.