Recommendation system is one of the modern applications to solve information overload problem in a way to provide recommendations of interest to users. Websites such as Amazon, Netflix, Facebook, YouTube, and others apply recommendation system in recommending their products. This also includes recommending news to the readers. However, the systems suffer from some challenges such as high dimensional data, data sparsity, and cold start. To address these problems, deep learning techniques have recently been integrated with recommendation systems and achieve good performance. Autoencoder is one of the most widely used deep learning techniques in recommender systems, especially used for feature extraction, data dimensionality reduction, fast convergence, unsupervised learning, and data reconstruction. In this paper, a performance comparison between three different models of autoencoder is presented which applying in the recommendation systems to further improve the quality of recommendations provided to users. The models are Hybrid Collaborative Recommendation via Semi-AutoEncoder (HRSA), Recommendation via DualAutoencoder (ReDa), and Hybrid Collaborative Recommendation method via Dual-Autoencoder (HCRDa). These models work by retrieving the potential latent factors from the sparse rating matrix and predict the missing ratings. The performance is compared based on reconstruction loss that applies to the MovieLens 100K dataset with three different sets of training data: 70%, 80%, and 90%. As a result, it was found that the HCRDa is outperformed the other models in terms of reconstruction loss based on the RMSE evaluation metric and the use of side information in the model. Thus, it is the most effective technique in terms of enhancing the quality of user recommendations.