Designing recommender systems to depolarize

Polarization is implicated in the erosion of democracy and progression to violence, which makes polarization properties large algorithmic content selection systems (recommender systems) a matter concern for peace security. While algorithm-driven social media do not seem be primary driver at country level, they could useful intervention point polarized societies. This paper examines depolarization interventions aimed transforming conflict: suppressing or eliminating conflict, but making it more constructive. Algorithmic considered three stages: what available (moderation), how selected personalized (ranking), presentation controls (user interface). Empirical studies online conflict suggest that only exposure-diversity proposed as an antidote ‘filter bubbles’ improved: under some conditions, can even worsen polarization. Using civility metrics conjunction with diversity may effective. However, diversity-based have been tested scale, work diverse dynamic contexts real platforms. Instead, intervening platform dynamics will likely require continuous monitoring metrics, such widely used ‘feeling thermometer’. These evaluate product features, potentially engineered objectives. using any metric optimization target harmful consequences, prevent processes from creating side effect prove necessary include measures objective function recommender algorithms.