Personalized recommendation has been receiving great attention because of its huge commercial value, but recommendation is also regarded as a difficult task due to the sparsity and noise in the real data. To address these issues and boost the recommendation performance, we propose an effective framework called Variational Collaborative Generative Adversarial Network (VCGAN), which improves the correlation between generated samples and real-world data by using the GAN-based structure. The generative network predicts the scores of the user's top-N items based on her profile and behavior, and the discriminative model aims to distinguish the predicted scores and maximize the objective function by learning from the ground truth. To avoid the high dimensionality problem, VCGAN adopts Auto-encoder (AE) to produce the latent vector of side information and performs adversarial training under the condition that using Variational Auto-encoder (VAE) as the generator. Without the high dimensionality problem, side information can be leveraged in the deep model to solve rating sparsity. Moreover, discovering richer latent representations through VAE helps the generator capture the distribution of real data more precisely. To evaluate the performance of our proposed model, we conduct extensive experiments on four real-world datasets against various baselines. The analysis and experimental results show the superiority of VCGAN over the compared methods.