The next paper in this WebSci 2016 session is presented by Mariana Arantes, whose interest is in the matching of video ads to YouTube videos. Such ads are displayed before some YouTube videos, and they can often be stopped after a set number of seconds. How do users consume these ads? How does their popularity change over time? What is the relationship between videos and ads, and does a better content match mean that ads are more likely to be watched all the way through?
It is difficult to investigate this given the poor range of data provided by the YouTube API, but this study was able to draw on Internet traffic data from a major Brazilian university to create a more comprehensive study of video consumption patterns. It identified 99,000 video ad views, for 58,000 unique videos and 5,000 unique video ads. The length of these ads varied widely; 31% ran for more than 2 minutes, and some even ran as long as two hours. Some 29% of ads were streamed in full; 35% were skipped within under 6 seconds, and 25% after more than 10 seconds.
Most ads were shown only a handful of times; only 3% were shown more than 800 times. The exhibition of most video ads was concentrated on a few days only. By examining the typical shape of the ad visibility curve over time, a number of typical such curves could be identified; very brief to brief visibility periods dominated, and these curve shapes were correlated to some extent wth different ad contents (e.g. music videos, entertainment ads, etc.). There is also a strong correlation between video ad and video content popularity.
Interestingly, video ad and video contents tend not to be paired very well, judging by the keywords associated with each of them; most video/ad pairings appear to be relatively random. There is great scope for a further improvement in such pairings, also drawing on user experiments.
