Click-through-rate Prediction Model In Text Search Based On Attention Mechanism

Click-through rate prediction(CTR)is a core issue in application scenarios such as recommendation system and search,and it is also an important indicator of the effectiveness of Internet advertising.Modern Internet advertising is mainly divided into search advertising and display advertising,of which search advertising is the largest form of advertising.Taking a search scenario as an example,a user searches for keywords to find a specified target,and by obtaining keywords with a commercial value intention and combining the results of click-through rate prediction,an advertiser displays candidate advertisements possibly clicked by users.A good CTR prediction model can accurately meet the needs of users and bring a good user experience,while allowing advertisements to maximize revenue.Through researching on click-through rate prediction models based on deep learning in recent years,there exists the fact that many existing click-through rate prediction models are insufficient in mining the cross-relationship between features,and the reuse of features is not high.Under the background,combined word embedding and attention mechanism,the model of deep convolutional network based on self-attention mechanism is designed.This problem is solved by designing reuse layers and deep convolutional networks.The main research work is as follows:Feature constructing was firstly performs on the real historical search data set,convert the text features into numerical features.Combining the advantages of XGBoost model with strong memory ability and Deep AFM model generalization ability,finally the fused click rate prediction model was designed.The fusion model does not consider the context information,and it ignores the dependence relationship between words.Based on this,combing with word embedding,a click-through rate prediction model based on hierarchical attention mechanism was designed.A two-layer bidirectional GRU network structure is used in the model,and the Attention mechanism is used for each layer.Based on the hierarchical attention model,a network structure that fully exploits the connection of semantic vectors in different dimensions is designed—Self-Attention-Based DeepConvolution Network(SDCN).Using AUC and accuracy as evaluation indicators,comparative experiments are conducted on the basis of the three models proposed in this thesis.From the experimental results,the AUC of SDCN model reaches 0.8477,which is 0.63%-2.14% improvement compared to other models.The accuracy rate is improved by 0.75%-2.06%.In the validity verification of the model substructure,the sub model fusion in the fusion model,the two-layer attention structure in the hierarchical attention model,and the reuse layer and outer product in the SDCN network Both convolutions have promoted the overall model.Taking the SDCN model as an example,the operation of the outer product has the largest promotion effect on the entire model.The convolution calculation is also essentially attention.A form of mechanism still has a promoting effect on the model.