Computational Multimedia Advertisement

Abstract:

Online advertising plays a crucial role in the online market. It remains the major source of revenue for many websites, and the important channel of promotional message broadcasting for many advertisers. Over the past decades, online advertising has been described as an economic issue, where a publisher uses auction models to sell online user's page view to advertisers. However, as the central to the advertising ecosystem, user's engagement on the displayed ads has been rarely considered. The market statistics reveal that the pervasive ads become annoying. Users tend to ignore the ads, or actively use ad-blocker tools to get rid of intrusive ads. The improperly displayed ads ruin user's online experience, waste advertiser's budget and further reduce publisher's future page visit. In this thesis, we aim to integrate multimedia techniques in online advertising to improve its effectiveness, so as to foster a healthy and vibrant online advertising ecosystem.

First, we propose an intelligent contextual video advertising approach, which utilizes contextual relevance, visual saliency and image memorability, to automatically associate the most suitable ad with an online video. The experimental results of an eye-tracking experiment demonstrated the effectiveness of newly introduced multimedia metrics. The visual saliency can increase user's eye-gaze duration on the displayed ads, and the image memorability can increase user's future ad recall rate.

Second, we propose a two-stage computational framework to integrate multimedia metrics to real-time bidding. To the best of our knowledge, we are the first to combine multimedia techniques with auction theory in the ad selection. We validated the proposed framework on several datasets, including ad auction datasets and multimedia datasets. Our experimental results show that the publisher can significantly improve the other stakeholders' benefits by slightly reducing her revenue in the short-term.

Third, we propose a novel monetization strategy that displays online ads in live streaming platforms. Given a live stream, we first embed a deep neural network to determine whether the current moment is proper to display an ad using the historical streams. Then, we detect a set of candidate ad insertion areas by incorporating image saliency, background map, and location priorities, so that the ad is displayed over the least important area. We introduce three types of relevance metrics including textual relevance, global visual relevance and local visual relevance to select the contextually relevant ad. To minimize user intrusiveness, we initially display the ad at a non-important area. If the user is interested in the ad, we will show the ad in an overlaid window with a translucent background. Empirical evaluation on a real-world dataset demonstrates that our proposed framework is able to intelligently display ads for live streaming videos while maintaining users' online experience.

To sum up, the three works have explored multimedia enhanced advertising strategies under different application scenarios, including contextual video advertising, display advertising, and live stream advertising. Extensive experiments on real-world datasets suggest that integrating multimedia techniques in online advertising can increase the quality of online advertising services. In the end, this thesis concludes the findings and discusses the future directions for computational multimedia advertising research.