Combined Effect of Content Quality and Social Ties on User Engagement: Research of Aesthetics Empowered by Algorithms
Rossano Schifanella, an Assistant Professor in Computer Science at the University of Turin, with experience in working in the industry (Yahoo Lab, Nokia Bell Labs) gave a talk about his research on social media data aiming at better understanding online as well as offline social behavior. His lecture covered several lines of research on the design of innovative methods for exploring urban spaces, related to the investigation of aesthetic evaluations in social media platforms. His research tried to answer the question which urban elements make people (un)happy, in order to provide the tools enabling us to find the shortest path between two locations which also takes into consideration how “pleasant” the path is (e.g. the smell, sights, and sounds of the street). The evaluation of an object’s pleasantness or beauty in general falls under computational aesthetics, which tries to algorithmically quantify the beauty of the pictures. This field of study is multidisciplinary, including inputs from philosophy, psychology, and neuroscience. The evaluations of aesthetic features of an object by a large number of people tend to converge, and Schifanella, with his colleagues, did a crowdsourcing experiment in which participants judged the beauty of pictures of different kinds of objects (nature, animals, people, urban spaces, see Figure 1) providing the “ground truth” data.
Using different machine learning procedures, researchers could quantify the beauty of a picture/photo. While popularity (sometimes called success in this research field) on social media is relatively easier to measure, the quality (referred to as performance – the “beauty” picture in this research context) is a more challenging concept to operationalize. Indeed, as he showed us through his case study, the two are not necessarily meaningfully connected on an individual level. For instance, two pictures on a certain image-sharing platform can be similarly aesthetically pleasant, but have different related user activities (favorites, comments, views) – so although they have similar quality, their popularity can be quite different. The analysis of the Flickr database (consisting of more than 15 billion pictures, 40 million users, and spanning over more than 10 years) showed that it follows a power law at user level: beauty ratings were distributed unequally, but not as much as popularity indices (e.g. favorites – only 2% of pictures had 6 or more likes). Based on their ground truth data, researchers could “discover” pictures that were not popular, although they would score highly on the quality measure.
While the relationship between social network and popularity, and popularity and quality have been studied, the relationship between social networks and quality has not been previously investigated. So, Schifanella and his colleagues tried to answer the question how social ties affect the quality consumption and production of this specific platform’s material? Firstly, according to their findings, quality and popularity indices are weekly connected (correlations in the range of 0.17 – 0.24). Secondly, the Majority Illusion seems to be at work – a tendency of users to see their neighbors as better than themselves, possibly because users can see only the local system. Thirdly, assortative mixing exists – those with similar quality are linked (Spearman’s rho is 0.48). Finally, the researchers wanted to learn more about the interplay between social ties and quality impact. They used matching experiments design to infer causality from observational data and concluded that in the short run, a new tie with a high-quality user is associated with the increase in the user’s quality. Regarding the long-term effects of quality, we will have to stay tuned for Rossano’s future research.
Blog post by Srebrenka Letina