Training A.I. to Recommend Ideal Preferences
Update: The paper has finally been published here. You can download it directly here.
Artificial intelligence is increasingly tasked with providing people with customized recommendations that are personalized for them. However, which version of the self should these recommendations be tailored for? The idea of a consistent self who retains the same preferences across contexts and time has been challenged by decades of work in consumer decision theory and behavioral economics. For instance, people’s preferences for products often differ from what they would like their preferences to be. Thus, programmers who develop recommendation agents face a challenge— whose preferences should the computer’s recommendations serve, the aspirational/ought self (who prefers high-brow documentaries) or the consuming/actual self (who prefers low-brow slapstick comedies)?
I collected training data of 1,000 people’s actual and ideal preferences for reading 52 different news articles based on a tweet/headline about each one. Then I built two machine learning models (using Random Forest) to 1) predict how much a given person would actually like to read each article and 2) predict how much a given person would ideally like to read each article - trained on their reading behavior for a subset of the other articles. I used Flask and Python to deploy these models to a server so that they could be run through a simple Web Request API.
In an ongoing follow up experiment, new participants answered questions about their reading behavior for the subset of “input” articles and were shown a recommendation score to a final article. Participants were randomly assigned to one of two conditions. In the actual preference condition, the score for the final article was the output of the model predicting their actual preference. In the ideal preference condition, the score for the final article was the output of the model predicting their ideal preference. Participants rated the accuracy of the recommendation, their satisfaction with the recommendation, their liklihood of seeking another recommendation, and chose whether or not to read the article.