PRIMA 2015, a really nice conference on Multi-Agent-Systems. My talk will focus on my past work on social influence in networks and processes of opinion polarisation and will discuss implications for the design of personalised recommender systems.
, FC, Italy
Personalization dramatically changed the Internet. Search engines provide results tailored to the interests of each individual user. Online markets recommend products based on the purchases of other customers who bought similar products in the past. Social networks rank incoming messages according to users’ interests. Personalization is of great help for users and is a multibillion-dollar business area. However, pundits warn that personalization creates cocoons of like-minded users, which makes the Internet boring and uninspiring. More worryingly, however, it has been warned that exposing users to ideas, news, and information that support their views will reinforce their opinions and, thus, foster the polarization of political opinions. These warnings received increasing attention, as opinion polarization might endanger societal cohesion and pose a challenge for political decision-making, as it impedes political consensus formation also on non-controversial issues. Reviewing the literature on social influence in networks and the conditions of opinion polarization, I will demonstrate in this talk that state-of-the art theory leaves us with great uncertainty about the consequences of personalization. In fact, two highly accepted models of opinion dynamics make opposing predictions about the consequences of personalization: Persuasion models, on the one hand, predict that personalization will increase polarization. Rejection models, on the other hand, imply that personalization will foster consensus rather than polarization. There is, thus, a pressing need to clarify which model better captures the effects of personalization. Second, I will describe the design of controlled experiments conducted on online social networks that allow calibrating the agents of existing influence models, which will make it possible to derive reliable predictions about the consequences of web personalization. Third, I will discuss implications of social-influence models for the development of personalized recommender systems. I will sketch different approaches to developing systems that generate personalized outcomes without fostering opinion polarization. I will show that such systems cannot be developed without an accurate model of social influence. On a more general level, I will conclude that the development of technologies on the Internet that have the potential to affect societal dynamics should be guided by theoretical and empirical research. In models of complex systems, even small and seemingly innocent differences in the assumptions about the underlying micro-mechanisms can have critical effects on macro-outcomes. As information technology affects micro-mechanisms, it crucial to understand possible consequences before it is too late to intervene.