The Empirical Study of Effects of Image Components in Ad Creative to Click by Computer Vision

Abstract

Production of components in ad creative itself is largely dependent on experience of designer. When you make creatives, it is possible to improve production process more efficiently if there is criteria or standard that guides you what components are more important. In this study, we conducted an empirical analysis to measure contribution of components in ad creative in the framework of CTR prediction for mobile advertising. By computer vision technology, human-interpretable keywords and color information were extracted as components that configure ad creative. As learner for CTR prediction, we identified which components in ad creative are effective for click by estimating importance and interaction of each feature value from the result of GBDT (Gradient Boosted Decision Trees) using a decision tree as weak classifier. The estimation of interaction was based on Interpretable Trees (inTrees) by Deng (2019). By combining computer vision technology and machine learning methods that can estimate importance and interaction of feature value, it enables not only to measure components of ad creative and their contributions, but also to expect wide range of applications.