<p>Food recommendation systems (FRSs) provide personalized food recommendations to users based on their taste preferences. In FRSs, users’ unique taste preferences are influenced by personal history, specific dietary needs, and the impact of social connection recommendations. These preferences are not captured efficiently in existing traditional FRSs because these preferences are quite complex. Therefore, deep neural networks (DNNs) came into existence to deal with this limitation. Deep neural networks dominate traditional recommendation techniques due to their ability to capture complex interactions effectively. This paper presents a novel approach known as "taste-centered deep matrix factorization model" to understand users’ food preferences through ratings. It hybridizes DNNs with matrix factorization for prediction and recommendation. In this paper, firstly, deep matrix factorization (DeepMF) utilizes a multi-layer perceptron (MLP) to handle the interactions between user and item along with their latent features. This entails feeding the suitable data into the MLP layers to learn intricate patterns and relationships. Secondly, the learned representations from the MLP are employed to generate appropriate food recommendations to users. This model assists users in developing good eating habits that align with their food choices and requirements. Experimental analysis on two standard datasets demonstrates that our proposed FRSs model outperforms several state-of-the-art models in terms of accuracy.</p>