Abstract: In the recommender system, deep learning models such as FiBiNET and AFM can focus on the importance of features for click-through rate prediction. FiBiNET's deep model uses DNN network to model feature interaction quite implicitly, but using DNN to learn higher-order features may lead to dilution of lower-order feature crossing. High-order important features were learned by superposing multiple SENET attention mechanisms and high-order attentive factorization machine was added to update feature representations. A new click rate prediction model HoFiBiAFM was formed. By comparing the classification task and regression task with other CTR prediction models on the Movielens-100K and Movielens-1M datasets, HoFiBiAFM's click-through prediction performance was verified.