Research On Advertising Click-through Rate Estimation Algorithm Based On Deep Learning

Advertising,as the most effective business profit mode of the current internet enterprises,has been gradually integrated into people’s lives under the continuous promotion of academia and industry.Real-time bidding advertising has become a widely used commercial delivery model because of personalized recommendations based on the different preferences of each user.And,the precise prediction of the Click-Through Rate can optimize the ranking of candidate advertisements,which directly affects the user’s experience,the effectiveness of advertisements,and the revenue of the platform.Therefore,an accurate CTR prediction algorithm is one of the core parts of the advertising system.For this reason,the article conducts further researches from three aspects: imbalanced data,explicit construction of high-order crossing features,and representation of user’s interest behind data,and achieves the following innovative results:1.In the real world,the number of negative samples in the advertising domain is much larger than the number of positive samples,and deep models generally assume that the number of samples in each category in the training data is relatively balanced.How to resolve the contradiction between the two has become a research direction.Hence,an adaptively equilibrated loss function is proposed for the imbalance problem of class distribution.According to the observed random variation on the ratio of the samples’ number of each category in the mini-batch,the label’s information of the samples is used to adaptively adjust the loss ratio of each sample in the loss function so that the total loss generated by samples from one category remains same order with the others.In addition,the idea of Online Hard Example Mining is also introduced.The experimental results on the unbalanced CTR datasets and the image multi-classification datasets show that the loss function can solve the unbalanced class distribution problem to a certain extent.2.In order to make up for the lack of crossing features in deep neural networks,the existing model concatenates embedding vectors together to construct high-order crossing features in the form of matrix mapping,but this obscures the relationship of feature fields.Hence,a new module for explicitly constructing high-order cross-feature information is proposed and then is combined with a linear module and a deep module to end-to-end learn semantic information of different levels.By preserving the FM’s structure in under-layer and using the feature vectors after crossing to recombine into new vector representations of sparse features,the steps of feature crossing and vector representation reconstruction of the sparse feature are repeated to obtain cross-feature information of different orders.The effects are fully evaluated and compared on three public datasets.Experiments show that the algorithm has certain advantages over other baseline models.3.User interest network,as a new model of Click-Through Rate prediction,has some disadvantages in using the inherent laws of data,and the problem of noise behaviors about users has not been solved well.Hence,an attention model that utilizes the user’s historical interaction behaviors from two different perspectives is proposed.In addition to the existing taking interacted item set of the user as the candidate user’s interest expression,the historical interacted user set of the item is also used as the characteristic expression of the candidate item.The attention mechanism is used to measure the differential preferences of the candidate user(item)on the historical interactive item(user)set and the semantic similarities between the candidate user(item)and the historical interactive user(item)set,thereby filtering out redundant noisy behaviors in the historical behavior list.Experimental comparisons on multiple challenging datasets verify the effectiveness and rationality of the method.