This paper introduces NCF (Neural Collaborative Filtering), a general framework that uses deep neural networks to model user–item interactions in collaborative filtering. Unlike traditional matrix factorization (MF), which relies on an inner product to model interactions, NCF replaces this with a neural architecture that can learn arbitrary functions from data. The framework is designed to handle implicit feedback, which is indirect and noisy, and is more common in real-world applications than explicit feedback. NCF is shown to be more expressive and flexible, as it can generalize MF and incorporate non-linear interactions through multi-layer perceptrons (MLPs). The authors propose two key models under NCF: GMF (Generalized Matrix Factorization), which is a linear variant of MF, and MLP, which uses non-linear interactions. They also introduce NeuMF, a hybrid model that combines GMF and MLP to leverage both linear and non-linear interactions. Extensive experiments on real-world datasets demonstrate that NCF outperforms state-of-the-art methods, with deeper networks yielding better performance. The paper also highlights the effectiveness of using a probabilistic approach with log loss and negative sampling for training, and shows that pre-training improves model performance. Overall, the study presents a promising direction for deep learning in recommendation systems.This paper introduces NCF (Neural Collaborative Filtering), a general framework that uses deep neural networks to model user–item interactions in collaborative filtering. Unlike traditional matrix factorization (MF), which relies on an inner product to model interactions, NCF replaces this with a neural architecture that can learn arbitrary functions from data. The framework is designed to handle implicit feedback, which is indirect and noisy, and is more common in real-world applications than explicit feedback. NCF is shown to be more expressive and flexible, as it can generalize MF and incorporate non-linear interactions through multi-layer perceptrons (MLPs). The authors propose two key models under NCF: GMF (Generalized Matrix Factorization), which is a linear variant of MF, and MLP, which uses non-linear interactions. They also introduce NeuMF, a hybrid model that combines GMF and MLP to leverage both linear and non-linear interactions. Extensive experiments on real-world datasets demonstrate that NCF outperforms state-of-the-art methods, with deeper networks yielding better performance. The paper also highlights the effectiveness of using a probabilistic approach with log loss and negative sampling for training, and shows that pre-training improves model performance. Overall, the study presents a promising direction for deep learning in recommendation systems.