Causally estimating the effect of YouTube's recommender system using counterfactual bots

Area of Science:

• Social Sciences
• Computer Science
• Media Studies

Background:

• Concerns exist regarding recommendation algorithms amplifying problematic and radicalizing content on online platforms.
• Evaluating algorithmic effects is challenging due to the lack of counterfactuals—what users would view without algorithmic recommendations.

Purpose of the Study:

• To causally estimate the role of YouTube's recommendation algorithms in the consumption of highly partisan content.
• To disentangle the effects of algorithms from user intentions in content consumption.

Main Methods:

• Introduction of a novel method termed 'counterfactual bots'.
• Comparison of bots replicating real user consumption patterns with counterfactual bots following rule-based trajectories.
• Analysis of YouTube's sidebar and homepage recommendation dynamics.

Main Results:

• Exclusive reliance on YouTube's recommender system leads to less partisan content consumption on average.
• This moderating effect is most significant for heavy partisan consumers.
• YouTube's sidebar recommendations adapt quickly, forgetting partisan preferences within approximately 30 videos, while homepage recommendations shift more gradually.

Conclusions:

• Individual consumption patterns on YouTube primarily reflect user preferences, especially since the 2019 algorithm changes.
• Algorithmic recommendations appear to play a moderating, rather than amplifying, role in content consumption.
• The platform's recommendation system demonstrates a capacity for gradual adaptation to user-initiated shifts in content preference.