Click-Through Rate (CTR) prediction is a fundamental task in personalized advertising and recommendation systems. This study proposes a model named ADditional Variational AutoEncoder (ADVAE) based on an improved denoising autoencoderto improve CTR prediction and cold-start. It adds random Gaussian noise to the input data and generates new embedded features by the improved denoising autoencoder. Then, multi-level features interact to predict the users’ clicking. This method can learn the relationship between feature embedding and interactions in data sparse and cold-start situations. In addition, it has strong robustness since it focuses on the interaction in feature domains and dynamically repairs feature embedding of low-frequency data. Besides, this method can be dynamically applied to other deep learning models, with high flexibility. The results show that the proposed approach outperforms its counter parts in terms of CTR prediction and cold-start.