| In recommendation systems,click-through rate prediction is a critical task,the purpose of which is to predict the number of clicks on advertisements in order to achieve effective recommendations to users and thus increase the revenue of online advertisements.In this paper,by analyzing the existing deep click-through rate prediction models,we find that the existing models do not consider the strength of the relationship between distinguishing features and interaction features,and have limitations in feature representation.To address the above problems,this paper proposes a Field-aware Attentional Embedding Deep Crossing Network(FAEDCN)model based on the attention mechanism.First,in the embedding part,the self-attention mechanism is combined with the field-aware embedding technique to weight attention to the embedded features;second,in the feature interaction part,the display interaction feature information is captured by the attentional cross-network module and the implicit interaction feature information is captured by the deep product network module;finally,the two modules are trained in parallel to output the final prediction value,which enriches the model representation and improves the model prediction at the same time accuracy of the model.In this paper,the FAEDCN model constructed in this paper is trained and tested on two major Internet ad click competition datasets of Criteo and Avazu,and the AUC index and Logloss index are used as the evaluation criteria for the model performance.The experimental results show that the FAEDCN model has higher performance compared with other models,and its AUC values increase by 1.40 percentage points and 1.56 percentage points,respectively,and its Logloss values decrease by 1.76% and 1.73%,respectively,compared with the DCN model.The paper has 36 figures,9 tables,and 59 references. |
