Abstract: The neural network-enhanced factorization machine (FM) model, which can capture more high-order feature interactions and improve the accuracy of predictions, has become a research hotspot in the field of recommendation algorithms. Aiming at the problem that existing models do not comprehensively consider high-order interaction features and original low-order features when modeling the interactions between users and items, and in order to improve the models ability to model user preferences, this paper proposes a new deep neural network-enhanced pairwise bilinear factorization machine model, DeepPRBFM, by combining depth neural networks and pairwise learning. This model adopts a bilinear structure with a pair of inputs containing positive and negative samples, utilizes multi-layer ResNet to preserve low-order features, enhances the interaction of high-order features with DNN, and employs a pairwise ranking-based loss function. Moreover, in the bilinear structure, increasing the proportion of negative samples can not only significantly alleviate the cold start problem of the recommendation system but also improve the prediction performance of the model. Experiments conducted on two real-world datasets show that the proposed model achieves higher recommendation accuracy and outperforms other models in objective metrics such as HR and NDCG.

Key words: implicit feedback, pairwise learning, factorization machine, neural network, cold start ,