Towards an interpretable advertisement click prediction

Abstract

In the new era of computational advertising, Click prediction models are used to anticipate a click response and guide marketers’ decisions about whom to target and how to personalize.The more prediction tasks achieve impressive performances, the more trust is put in black models to make important decisions in a business domain. Due to the complexity and lack of transparency of the models, posterior explanation methods are needed to identify features’contributions that envision a global explanation of the model. This thesis develops anadvertisement Click prediction using a supervised machine learning framework and uses theKernelSHAP method to provide feature importance insights on the predictions made by the model. The thesis aims to answer: 1) how can marketers integrate Click prediction in their businesses; 2) which are the most impactful features for a click response; 3) how advertisementcategories differ from an overall model. For that purpose, is used a publicly available ADS dataset (2016) to train a neural network. The results showed the overall model performance is not substantially different from a segmented categories performance. The output of KernelSHAP showed that even though the visual content is impactful for the likelihood of a Click for all models, each category has its own feature importance pattern influenced by the product the category promotes. The performance metrics presented a high ratio of true negativeto true positive due to a class imbalance problem. To mitigate the cost of misclassification is suggested an individual analysis that better fit target business model.

Date of Award	27 Jan 2023
Original language	English
Awarding Institution	Universidade Católica Portuguesa
Supervisor	Ana Guedes (Supervisor)