Predicting Visual Fixation on Digital Advertisement using Machine Learning

Sammanfattning: Despite the fact that ad-tech being a multi-billion dollar industry, the percentage of digital ads which are actually being clicked on is as low as 0.1 % in many cases. The performance of ads which can be clicked on are often measured by click-through rate (CTR), in other words, action based. However, for ads in which the advertiser’s goal is to only evoke brand/product awareness, they have to rely on metrics such as impressions or in-views. This master’s thesis aims to investigate if advertisement fixation can be used as a complementary metric to CTR. This was formulated by two research questions. The first one focused on to what extent machine learning models can predict advertisement fixation using web browsing data including eye tracking observations from a passive panel in Sweden, whereas the second research question focused on to give evidence into the most prominent features when predicting advertisement fixation. Predictive performance of four popular machine learning models, previously used in CTR prediction; Logistic Regression, Random Forest, XGBoost, and Field-Aware Factorization Machines were analyzed. Logistic Regression and Random Forest along with k-fold cross validation were used to validate the process of incremental feature engineering. The results demonstrated that an ensemble of three of the models could predict advertisement fixation with an F1-score of 0.5972, and an AUC-ROC value of 0.8005, where the latter is comparable to previous research in CTR prediction. In addition, the most prominent features when predicting advertisement fixation were concluded to be hostname, brand, ad type, x-coordinate, as well as the width and height of the ad. In conclusion, this shows that advertisement fixation can be predicted based on web browsing data. Further research is needed to determine if advertisement fixation should be used as a complement to CTR, and whether it will be adopted by the ad-tech industry.