Neural Collaborative Filtering

The paper tackles collaborative filtering-the key problem in recommendation-using deep neural networks on implicit feedback, an area that had received relatively little scrutiny compared with deep learning's success in speech recognition, computer vision, and natural language processing. The authors observe that prior deep-learning recommenders mostly applied neural networks to auxiliary information, such as textual item descriptions or acoustic features of music, while still modeling the crucial user-item interaction with matrix factorization and an inner product over latent features. Their framework, NCF, replaces that inner product with a neural architecture that can learn an arbitrary interaction function from data and uses a multi-layer perceptron to introduce non-linearities.

NCF is presented as a generic framework that can express and generalize matrix factorization as a special case. Extensive experiments on two real-world datasets show significant improvements over state-of-the-art methods, and the empirical evidence that deeper neural networks yield better recommendation performance underscored the value of learning user-item interactions with neural networks rather than fixed inner products.