Research On Click-Through Rate Prediction Based On Feature Interaction

Click-through rate prediction is a measure of how likely a user is to click on a particular advertisement to recommend the appropriate advertisement.Advertising information contains a lot of information such as user characteristics,advertising characteristics,and context characteristics,which can be used to predict user click rate.Feature interaction can improve the accuracy of user click rate prediction using feature information.The development of an accurate and efficient click-through rate prediction model has become the focus of current research.However,the current click-through rate prediction model still has some shortcomings.On the one hand,it is based on the fully connected neural network model,which concatenates the embedding vectors and inputs the fully connected network,which cannot model explicit feature-level interactions.On the other hand,the existing model enumerates feature interactions,which cannot model the relative importance of adaptive feature interactions and adaptive feature interactions.In this thesis,two prediction models are constructed to solve the different problems in the existing click-through rate prediction,and the models are verified by experiments.The main work is as follows:(1)Research on the recurrent attention feature interactionTo improve the click-through rate prediction effect,the recurrent attention interactive click-through rate prediction model RAt I was constructed by mining some correlation and vector-level feature interactions.Furthermore,combined with deep neural networks to mine explicit and implicit feature interactions,an integrated click-through rate prediction model RAt I+ is proposed for deep attention interaction.First,the click-through rate prediction model of RAt I takes the feature embedding vector as the basic operating unit,and uses the attention mechanism to model the feature interaction at the embedding vector level.This feature interaction mode realizes an explicit feature interaction mode,makes full use of the attention mechanism to model the correlation between features,and learns more relevant and meaningful feature interactions,so as to improve the click-through rate prediction effect.Secondly,by integrating the deep neural network and RAt I to jointly model implicit and explicit feature interactions,the modeling effect is further improved,and an integrated model of RAt I+ is constructed.Finally,experimental verification was carried out on the Movielens dataset and Frappe dataset,and the results were compared with some major baseline models to prove the effectiveness of the model.(2)Research on dual adaptive and hierarchical feature interaction click-through rate predictionBased on the attention logarithmic neural network,the adaptive feature interaction is mined,and the importance of the feature interaction is learned adaptively.A dual adaptive feature interactive click rate prediction network module ALN is constructed.Furthermore,a dual adaptive and hierarchical feature interactive click-through rate prediction model(ALHIN)was constructed by combining the hierarchical interaction module to improve the prediction effect while maintaining the efficiency of model.Firstly,the dual adaptive feature interactive click-through rate prediction network module uses the attention logarithm neural network.On the one hand,it learns the exponential parameters of the logarithm of attention neural network to model the feature interaction adaptively.On the other hand,it models the interactive importance of features adaptively through the attention mechanism.Secondly,the dual adaptive feature interaction click-through rate prediction model is based on ALN,combined with hierarchical interaction module to learn feature interaction layer by layer iteratively.The hierarchical interaction module iteratively learns the feature interaction layer by layer on the feature embedding vector.On the one hand,the noise effect of ALN on the absolute value operation of the embedded vector is reduced.On the other hand,while maintaining the efficiency of model training,it also improves the effect of model.Finally,experimental verification studies are carried out and compared with the results of the models of the main algorithm to verify the effectiveness of the proposed model.Furthermore,the rationality of each module in the proposed model is verified by the ablation experiment.