Are online recommendation algorithms polarising users’ views?

The num­ber of people hold­ing extreme views, on sub­jects such as polit­ics, reli­gion or cli­mate change – to cite just three examples – has increased in recent years¹²³. This “polar­isa­tion”, as it is called, is dan­ger­ous, as it could poten­tially weak­en demo­cracy itself if allowed to spread unhindered. Online plat­forms such as social media play an import­ant role in this con­text, but the mech­an­isms by which they foster polar­isa­tion are not yet fully understood.

“Recom­mend­a­tion algorithms pro­foundly shape our digit­al exper­i­ence today, determ­in­ing the films we watch or the songs we listen to,” explains Giord­ano De Mar­zo. These algorithms are widely used by most of the web­sites we vis­it every day, the best-known examples being the “sug­ges­ted for you” mes­sages on Face­book, the “recom­men­ded items” on Amazon or Google’s PageR­ank sys­tem. They are designed to give us easy access to the con­tent most likely to interest us, and max­im­ise our engage­ment with the platform.

This is a sig­ni­fic­ant step for­ward in cre­at­ing a more bal­anced and inclus­ive online inform­a­tion ecosystem

A team of research­ers led by Giord­ano De Mar­zo, from the Depart­ment of Phys­ics at the Uni­ver­sity Sapi­enza in Rome, Italy, has stud­ied how a col­lab­or­at­ive user-to-user fil­ter­ing algorithm affects the beha­viour of a group of people repeatedly exposed to it. This type of recom­mend­a­tion algorithm is routinely used by online retail giants such as Amazon to identi­fy new con­tent, based on past activ­ity, that will be of most interest to users. Using ana­lyt­ic­al and numer­ic­al tech­niques, the research­ers were able to sim­u­late how the users’ con­tent pref­er­ences change in response to algorithmic recom­mend­a­tions. Their ana­lyses revealed three dis­tinct regimes or ‘phases’ in the user-base’s state that trap people in so-called “fil­ter bubbles”.

These states depend on key factors such as the “strength” with which the algorithm recom­mends items that are liked by sim­il­ar users, or that are pop­u­lar over­all. The study also iden­ti­fied strategies that allow an algorithm to provide per­son­al­ised recom­mend­a­tions without cre­at­ing fil­ter bubbles. This could con­trib­ute to the devel­op­ment of less polar­ising algorithms in the future.

Col­lab­or­at­ive fil­ter­ing⁴⁵ is one of the best-known and most widely used recom­mend­a­tion algorithms. It relies on the prin­ciple that the past beha­viour of users can be exploited to identi­fy new con­tent that they will enjoy the most. The down-side is that these algorithms can lead to a feed­back loop. This loop nat­ur­ally tends to bias future choices, redu­cing the diversity of con­tent avail­able. It is a loop of this kind that leads to fil­ter bubble effects, where users are not exposed to new or dif­fer­ing per­spect­ives, but simply to news and con­tent aligned with their exist­ing beliefs. In short, these loops con­trib­ute to “polar­isa­tion” . They are sim­il­ar to “echo cham­bers” , which have been more widely stud­ied⁶⁷⁸. How­ever, the dif­fer­ence is that bubbles are pro­duced by algorith­mic­ally-biased recom­mend­a­tions on online plat­forms, rather than from inter­ac­tion between like-minded users.

In this new study, pub­lished in Phys­ic­al Review E, Giord­ano De Mar­zo and his col­leagues found that depend­ing on two para­met­ers, the strength of the sim­il­ar­ity bias (a) and the strength of the pop­ular­ity bias (b), a col­lab­or­at­ive fil­ter­ing sys­tem can exist in three dif­fer­ent phases. These are dis­order, con­sensus and polar­isa­tion. Fur­ther­more, when both biases are suf­fi­ciently strong, the sys­tem forms polar­ised groups, lead­ing to the “fil­ter bubble” effect. For­tu­nately, this dis­ad­vant­age can be avoided at the bound­ary between dis­order and polar­isa­tion. Indeed, an algorithm at this bound­ary can provide mean­ing­ful recom­mend­a­tions without indu­cing opin­ion polar­isa­tion or trap­ping fil­ter bubbles.

“Our research provides a sys­tem­at­ic approach to quan­ti­fy­ing and ana­lys­ing the impact of col­lab­or­at­ive user-user fil­ter­ing,” explains Giord­ano De Mar­zo. “By employ­ing a stat­ist­ic­al phys­ics approach, we were able to sim­u­late and ana­lyse how users’ con­tent pref­er­ences change in response to algorithmic recommendations.” 

The new meth­od relies on a com­bin­a­tion of math­em­at­ic­al mod­el­ling and com­puter sim­u­la­tions. “In par­tic­u­lar, we have exploited tech­niques such as stochast­ic pro­cesses the­ory, prob­ab­il­ity the­ory and polya urn mod­els (a fam­ily of urn mod­els that can be used to inter­pret many com­monly-employed stat­ist­ic­al mod­els). On the com­puter side, we lever­aged Monte Carlo sim­u­la­tions” , explains Giord­ano De Marzo.

These ana­lyses could con­trib­ute to the devel­op­ment of new meth­od­o­lo­gies for design­ing effect­ive recom­mend­a­tion algorithms, he adds. “By under­stand­ing the mech­an­isms that lead to ‘fil­ter bubbles’, we can devel­op sys­tems that favour a wide range of con­tent, thereby mit­ig­at­ing the risks of polar­isa­tion while enhan­cing user engage­ment and con­tent diversity. This is a sig­ni­fic­ant step for­ward in cre­at­ing a more bal­anced and inclus­ive online inform­a­tion ecosystem” .

The research­ers will now study the impact of inter­ac­tions between users (as com­monly observed in social net­works) on recom­mend­a­tion algorithms. “Adding this para­met­er could con­sid­er­ably enrich our under­stand­ing of the inter­play between social dynam­ics and algorithm-driv­en con­tent dis­tri­bu­tion. This will provide a more hol­ist­ic view of digit­al envir­on­ments”, explains Giord­ano De Marzo.

They will also study the role of link recom­mend­a­tion algorithms, that is, those that sug­gest friends we will link to. Finally, they are cur­rently exploit­ing Large Lan­guage Mod­els for power­ing more real­ist­ic sim­u­la­tions. “These sim­u­la­tions will be the ideal start­ing point for a more detailed under­stand­ing of online dynam­ics and recom­mend­a­tion algorithms,” he concludes.

Isabelle Dumé