Signed graph representation learning is a hot research topic nowadays. Existing methods pay attention to existing links, largely overlooking the importance of non-existent links. Due to limitations of balance theory and sparsity in signed graphs, these models struggle to capture the complexity of real-world signed graphs, leading to poor performance in downstream tasks like link prediction. To address these issues, we propose a Multi-head Variational Signed Graph Auto-Encoder (MVSGAE). Specifically, we design multiple variational inference networks to extract the node representations from positive and negative subgraphs, which only contain the positive and negative edges from the original graph. Then, we reconstruct the corresponding adjacency matrices of both subgraphs using node representations. Finally, we use the predictions from two different subgraphs to decide whether a link exists and what its sign is. Experimental results show MVSGAE outperforms various competitive baselines, achieving state-of-the-art performance.