Learning Personalized Decision Support Policies

2023-04-13

arXiv.org

AI Summary

This study addresses the problem of dynamically matching cost-effective AI decision support modalities to heterogeneous users. We propose Modiste, the first framework for learning personalized decision support policies. Grounded in stochastic contextual bandits, Modiste learns optimal support strategies—such as expert consensus or large-model predictions—online without prior assumptions about user preferences. It supports multi-objective optimization (e.g., accuracy–cost trade-offs) and accommodates interactive learning across both vision and language tasks. Our key innovation lies in formulating decision support as a dynamic policy selection problem, enabling real-time, cross-tool and cross-modal adaptation. Empirical evaluation with human participants demonstrates that Modiste significantly outperforms static baselines: under cost-sensitive conditions, it reduces average support cost by 47% while sacrificing less than 2% in task performance—validating its practical efficacy and deployment viability.

Abstract

Individual human decision-makers may benefit from different forms of support to improve decision outcomes, but when each form of support will yield better outcomes? In this work, we posit that personalizing access to decision support tools can be an effective mechanism for instantiating the appropriate use of AI assistance. Specifically, we propose the general problem of learning a decision support policy that, for a given input, chooses which form of support to provide to decision-makers for whom we initially have no prior information. We develop $ exttt{Modiste}$, an interactive tool to learn personalized decision support policies. $ exttt{Modiste}$ leverages stochastic contextual bandit techniques to personalize a decision support policy for each decision-maker and supports extensions to the multi-objective setting to account for auxiliary objectives like the cost of support. We find that personalized policies outperform offline policies, and, in the cost-aware setting, reduce the incurred cost with minimal degradation to performance. Our experiments include various realistic forms of support (e.g., expert consensus and predictions from a large language model) on vision and language tasks. Our human subject experiments validate our computational experiments, demonstrating that personalization can yield benefits in practice for real users, who interact with $ exttt{Modiste}$.

Problem

• Decision Assistance
• Cost-effectiveness
• Individual Decision-making

Innovation

• Personalized Decision Support
• Multi-Armed Bandit Algorithm
• Cost-Effective Assistance