Multimodal Ad Recall Prediction Based on Viewer’s and Ad Features

Ad recall is a commonly used measure of advertising effectiveness. Automatic prediction of advertising effectiveness will help to improve video advertising and optimize the process of ad creation. We present a novel multimodal approach to ad recall prediction for video advertising based on viewer’s features and ad features. In our experiment twenty people watched ads (n=100 in total). Ads have ground truth ad recall that was previously obtained in a field study. While people were watching ads, we recorded them with video camera, collected contact photolpletysmography and eye-tracking data, and also asked them to complete questionnaires. From these data we extracted “viewer’s features” –emotional, physiological and behavioral parameters. As well, we had “ad features” – target ratingPoint (TRP) and weighted target rating point (WTRP) metrics. To predict ad recall from these features a range of regression models were tested. Random Gaussian projection with Support Vector Machines howed the best performance (MAE=0.09, R=0.6).