Research On Deep Image Information In Search Advertising

With the development of Internet technology,Internet advertising is in-creasingly welcomed by advertisers.As a new form of advertising,Internet advertising has gradually become the most important type of advertising for modern commercial advertising.Search advertising is the largest form of In-ternet advertising.Traditional search advertising is generally based on the NLP model method.However,in the field of Product Ads,the information of the text is often not enough to fully express the information of a product.This paper apply the image features and the framework of Teacher-Student into the Prod-uct Ads scene of search advertising.Specifically,there are the following two Aspects:1.Based on the scene of Product Ads,this paper proposes a series of methods and strategies to construct image features,and compares the effect of different image features on the final Query and Ads matching models to find the most effective image feature construction scheme.2.Apply the Teacher-Student architecture into the search advertising and proposing a new training scheme.First,use the annotated data to train multiple different Teacher-models for model fusion,then use the fused Teacher-model to calculate scores for both the annotation data and the log data,for the training of the Student-model,and finally improve Model effect.This paper mainly proposes improvements from the image features of ad-vertisements and the training schemes of online models,and has achieved sig-nificant improvement in the actual test set of Microsoft Bing,which has certain theoretical and application value.