1. médialab Sciences Po
  2. News
  3. How do recommendation systems learn our political opinions?

How do recommendation systems learn our political opinions?

In a study published in PLOS One, Tim Faverjon, Jean-Philippe Cointet and Pedro Ramaciotti demonstrate how recommendation systems can learn to represent their users' political orientations based on behavioral traces.

Post

Recommendation systems now play a central role in how we access information on social media. By selecting the content presented to users, they help shape their information environment and raise questions, in particular, about the diversity of the content to which users are exposed and the risks of political segregation.

But how does a recommendation system come to learn its users' political views? That is the question that Tim Faverjon (PhD student at the médialab), Jean-Philippe Cointet (researcher at the médialab) and Pedro Ramaciotti (researcher at the médialab) have been exploring in a study published in the scientific journal PLOS One.

To conduct this study, the three researchers designed an experiment using data from a panel of 40 000 X users, a randomly selected sample representative of the ideologically engaged core of the platform’s political sphere. A recommendation algorithm was then trained using the URL links shared on X by this panel over a 4-month period. This allowed them to identify the dimensions that shape the recommendations. Explanability methods were then used to measure the role of the various characteristics.

The results highlight the ability of recommendation systems to learn political orientations from users’ behavioral traces, without this information being explicitly provided to them. The study also shows that the way political dimensions are structured within the algorithm’s internal model can influence recommendations, as evidenced by how these recommendations change when the model ignores its political representations. These findings represent a step toward a better understanding of the implicit norm implemented by algorithms and how it operates.

Read the article: How do Recommender Systems Learn Political Opinions? A Semi-Synthetic Step-by-Step Experiment