Research On Advertising Click-through Rate Prediction Methods Based On Deep Learning

Advertising click-through rate prediction is an important task in online advertising push,improving its accuracy can not only bring users satisfactory advertising experience,but also bring huge economic benefits to the platform and advertisers.Although existing deep learning models have achieved good results,they still have some deficiencies:(1)The latent relationships between features are not sufficiently learned: Firstly,the feature interaction is realized directly in the form of vector product,ignoring the potential contexts between different feature interactions.Secondly,the way of feature interaction is unitary,so it is impossible to learn the different behavior relationship between the same feature pair.Finally,the attention mechanism is widely used in features of the same property,ignoring the weight relationship between feature representations of different properties.(2)The expression ability of feature information is insufficient and the interpretability of higher-order features is insufficient: Firstly,feature interaction is conducted on the embedded dimension.Because the embedding process of features is a simple linear operation and the interaction result of embedded features will directly affect the prediction effect of the whole model,the expression ability of the model is insufficient;Secondly,although the deep neural network has a strong fitting ability to learn higher-order information of features,it cannot make a reasonable explanation for the features of each order,so it learns an implicit higher-order relationship,which brings the problem of insufficient interpretability to the model To solve these problems,this paper proposes two novel deep learning models,and the specific research contents are summarized as follows:(1)In order to model and learn the potential diversity relationship between features,this paper proposes a novel deep learning model called Multi-Relational Interaction Network(MRIN).On the one hand,MRIN designs two feature interactive modules.The two modules use different multidimensional parameter matrices to fit the potential contexts of feature interactions and combine two different interaction modes to learn different interaction behaviors of the same feature pair.On the other hand,MRIN introduces the attention mechanism and learns the importance between the interactive features and the embedded features of the two feature interaction modules through three attention networks.In addition,MRIN also designs a global attention module to dynamically learn the weight relationship between the feature representations of the output of different attention networks.(2)In order to improve the information expression ability of features and the interpretability of higher-order features,this paper proposes a novel deep learning model called Combining Multi-Head Self-Attention and Channel Attention For CTR Prediction(MHSACA).Firstly,the model introduces a multi-head self-attention mechanism on the embedding dimension,and maps the original embedded feature matrix into multiple self-attention subspaces,so that the embedded feature matrix is aggregated to richer information and new feature matrix is generated.Then the feature matrices of finite order are generated by multiple feature crossings,and the channel attention mechanism is introduced to convolution the feature matrices of each order,and the explicit interpretation of each order feature matrix is further given.(3)In this paper,we conduct a large number of experiments on the two proposed models on two real and public datasets,Criteo and Avazu,including comparison with the baseline models,ablation experiments,and hyperparameter tuning.Experimental results prove the effectiveness of the work in this paper.