Investigation of Higher Order Features for Click Prediction in Image Ads

Abstract

Click prediction of image advertisement is an important task to create more effective ads and successful marketing campaigns. Many existing studies perform click prediction based on features extracted from input images by pre-trained networks to solve image classification tasks. Unlike general images, however, ad images contain more complex contextual information. Therefore, the extracted features may not have sufficient information for such purpose. In this study, we create contextual features obtained by human annotation and another pre-trained network to solve tasks that are oriented on human perception. We investigate how such features contribute to click prediction task by experiments using actual ad impression data.