A common method for creating recommender systems that can forecast users’ preferences for items based on their prior interactions with those items is collaborative filtering. Matrix factorization is one example of a traditional collaborative filtering technique that has difficulties processing sparse and high-dimensional data. In this article, we provide an innovative method for collaborative filtering that makes use of autoencoder neural networks. In particular, we learn user and item embeddings from the user-item interaction data using a deep autoencoder architecture. The missing ratings are then predicted using the learned embeddings and a rating prediction function. We also suggest a pre-training and fine-tuning technique to enhance the autoencoder’s functionality. Our test findings on the benchmark dataset demonstrate that the suggested strategy performs better than traditional collaborative filtering methods, including matrix factorization, and is computationally efficient in handling sparse data. The proposed method provides a promising direction for building more accurate and efficient recommender systems.