ELIXIR: Learning from User Feedback on Explanations to Improve Recommender Models

Explanations can improve recommendations, not just user trust: Rather than treating explanations as a one-way tool to help users understand recommendations, ELIXIR shows that user feedback on why a recommendation was made can actually be used to make better recommendations in the future—a finding that suggests explanations serve a dual purpose in responsible AI systems.

User feedback on explanation pairs outperforms traditional methods: The study demonstrates that asking users to compare pairs of recommendations and their explanations generates more useful learning signals than conventional item-level feedback alone, suggesting that the reasoning behind recommendations matters as much as the recommendations themselves.

Human-in-the-loop systems can overcome data scarcity: By incorporating label propagation with item-similarity neighborhoods, ELIXIR addresses a common challenge in personalized AI—the sparseness of user feedback—meaning systems can learn user preferences more effectively even with limited interaction data.

Real-world testing validates the approach: The framework was tested through an actual user study on movie and book recommendations, showing "significant improvements" over baseline methods, which demonstrates that this explanation-feedback approach works in practice, not just in theory.

Transparency and performance can align: This research suggests that building AI systems to be more transparent (through explanations) doesn't require sacrificing recommendation quality; instead, transparency mechanisms can actively contribute to better system performance.