Towards Autonomous Decision-Making : A Probabilistic Model for Learning Multi-User Preferences

Information systems have revolutionized the provisioning of decision-relevant information, and decision support tools have improved human decisions in many domains. Autonomous decision- making, on the other hand, remains hampered by systems’ inability to faithfully capture human preferences. We present a computational preference model that learns unobtrusively from limited data by pooling observations across like-minded users. Our model quantifies the certainty of its own predictions as input to autonomous decision-making tasks, and it infers probabilistic segments based on user choices in the process. We evaluate our model on real-world preference data collected on a commercial crowd sourcing platform, and we find that it outperforms both individual and population-level estimates in terms of predictive accuracy and the informativeness of its certainty estimates. Our work takes an important step toward systems that act autonomously on their users’ behalf