Research On Click-through Rate Prediction For Image Advertising Based On Deep Learning

With the flourishing development of internet and e-commerce,Online Display Advertising(DA)has become one of effective ways for businesses to promote their products by showing textual,image,or video ads on various web pages.Click-Through Rate(CTR)prediction of image ads plays an important role in the accurate delivery of display advertisements.On the one hand,the results of CTR prediction provide a quantitative basis for effectively improving user experience and enhancing their viscosity.On the other hand,CTR prediction is an important way to improve the effect of internet advertising marketing.Currently,most of the related work is based on textual and image features,and statistical learning methods are used to achieve CTR prediction.On this occasion,image features mainly rely on manual selection,which is a long haul and may not necessarily bring about performance improvements.Besides,the meaning of each feature is isolated and the interactions between features are not fully represented within the process of CTR prediction.Therefore,how to quickly and effectively screen features in advertising data and fully explore the interactions between features has become the key of further increasing performance.Recently,deep learning is widely used in automatic feature extraction and has achieved remarkable results.We carry out relevant research on CTR prediction based on deep learning technology,and propose an end-to-end prediction method for image advertising.Specifically,in this paper,an image ad has visual features and basic features,and the convolution neural network is employed to extract visual features from an ad image.After that,the deep neural network is used to explore the intrinsic interactions between features to give a better performance.Compared with hand-crafted image feature extraction,our method is more efficient and can fully explore the interactions to make more accurate results.We design and develop multiple experiments on the data set of a commercial advertising platform in 2017.The experimental results show that compared with other methods,our proposed method can effectively improve the RMSE and MAPE metrics with highly sparse inputs,and achieves the expected effect.