Figure 6.

Simulated impact of different recommender systems on users' satisfaction and content diversity exposure. Satisfaction is assumed as a proxy for the sustainability of the social media platform. Content diversity exposure could play an important role in countering the effects of filter bubbles (Bozdag and van den Hoven, 2015; Nikolov et al., 2015), echo chambers (Wolfowicz, 2015; Bessi, 2016; Gillani et al., 2018), and ultimately society polarization. Mean and standard deviation over 10 runs with three different new connections Recommenders: maximize opinion diversity, random, overlapping neighborhood. For each strategy (colors) Satisfaction (“o”) and diversity (“x”) are pictured. Overlapping: recommend users with the highest number of common friends. Diversified: recommend users with the highest opinion difference. Random: baseline, recommend random users. Satisfaction: the mean distance for each user between his opinion and the ones in his feed. Diversity: entropy of binned opinions that populate users' feed in each time step. Highlighted areas represent standard deviation across different runs. Each social network is initialized with 100 users (nodes) and connections (edges) are created with an adaption of preferential attachments (Albert and Barabási, 2002). Differently from Albert and Barabási (2002) the nodes' probability of being connected with an incoming node is not proportionally related to nodes' degree but is related with their opinion distance. Users were modeled by extending the model proposed in Geschke et al. (2019) with a backfiring component (Bail et al., 2018), i.e., users exposed to opinions that were distant from theirs moved in the opposite direction. The recommender that maximizes diversity between the pairs of users to connect showed a slower start but achieved higher exposition to more diverse content and a similar level of satisfaction to the other two RSs.