Recommender systems are essential in digital services for helping users find relevant items. One of the main challenges faced by these systems is the problem of sparsity, where limited user-item interaction data makes it difficult to generate accurate recommendations. Various conventional collaborative filtering such as Matrix Factorization (MF) and Singular Value Decomposition (SVD) have been developed to overcome the sparsity problem. However, these methods have limitations in capturing the non-linear relationships between users and items, thus falling short in effectively handling high levels of sparsity. Therefore, we proposes a Deep Matrix Factorization (DMF) model, which integrates deep learning techniques with matrix factorization to capture non linear patterns in user-item interactions. DMF utilizes a neural network structure to extract more complex latent representations and improve prediction accuracy under highly sparse data conditions. The model was tested on four datasets with varying levels of sparsity: Amazon Books Reviews, MovieLens Small Latest, MovieLens 100K, and Netflix Prize. The experimental results show that DMF consistently produces lower Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) values on all four datasets, with average decreases in MAE and RMSE of 11.81% and 12.23% on Amazon Books Reviews, 9.52% and 11.91% on MovieLens Small Latest, 5.83% and 6.11% on MovieLens 100K, and 6.41% and 6.64% on Netflix Prize, respectively, compared to MF and SVD methods. This demonstrates that integrating deep learning into recommender systems can significantly enhance performance, especially in addressing the challenges posed by sparsity.