Emilio Calvano (Università di Bologna)

Upon logging into their Netflix, Amazon or Spotify accounts, consumers are usually greeted with personalised recommendations about goods from the catalogue that they might enjoy or need. These recommendations are provided by highly sophisticated algorithms, called recommender systems, which use big data to predict consumer tastes. A known challenge for Recommender Systems is understanding when to not make a recommendation. The reason being that consumer confidence in these systems is built over time through past experiences and quickly evaporates after recommendation errors.
We propose a model capturing the recommender incentive to build a reputation for being accurate and show that in equilibrium it leads to biased product recommendations.
The theory delivers a number of predictions that can be tested in a controlled laboratory setting. We provide experimental evidence that recommendations matter in the sense that they affect our subjects' consumption choices. Secondly we document that consumers learn on the accuracy of the recommender from experience. Finally we test whether the model prescription enhances the recommender’s profit leading to suboptimal participation decisions.