Exploiting Preference Elicitation in Interactive and User-centered Algorithmic Recourse: An Initial Exploration

TL;DR

The paper tackles the need for actionable, user-tailored recourse in algorithmic decisions by proposing a guided-interaction framework that elicits user preferences via pairwise comparisons and updates a cost-weight distribution $P(W)$, integrated with reinforcement learning and Monte Carlo Tree Search to efficiently generate recourse plans. It contrasts this with a purely exploratory GAM Coach–style interface in a money-lending scenario to assess user experience across pragmatic and hedonic dimensions. Findings suggest guided interaction can enhance perceived efficiency and certain UX qualities when users spend more time, but may constrain exploratory behavior; conversely, exploratory interfaces offer more freedom yet risk perceived inefficiency. The work highlights the value of hybrid designs that support exploration while gently guiding users toward effective solutions, while acknowledging small-sample limitations and the need for broader validation in real-world settings.

Abstract

Algorithmic Recourse aims to provide actionable explanations, or recourse plans, to overturn potentially unfavourable decisions taken by automated machine learning models. In this paper, we propose an interaction paradigm based on a guided interaction pattern aimed at both eliciting the users' preferences and heading them toward effective recourse interventions. In a fictional task of money lending, we compare this approach with an exploratory interaction pattern based on a combination of alternative plans and the possibility of freely changing the configurations by the users themselves. Our results suggest that users may recognize that the guided interaction paradigm improves efficiency. However, they also feel less freedom to experiment with "what-if" scenarios. Nevertheless, the time spent on the purely exploratory interface tends to be perceived as a lack of efficiency, which reduces attractiveness, perspicuity, and dependability. Conversely, for the guided interface, more time on the interface seems to increase its attractiveness, perspicuity, and dependability while not impacting the perceived efficiency. That might suggest that this type of interfaces should combine these two approaches by trying to support exploratory behavior while gently pushing toward a guided effective solution.

Figures (3)

• Figure 1: The guided interface that uses the algorithmic recourse method enriched with preference elicitation: the user can only express preferences on the proposed features since they have been selected by the algorithm as those that optimize the learning of a personalized cost function.
• Figure 2: Control interface used in the study; inspired by GAM Coach wang_gam_2023: the system proposes one or more plans, and the users can express their attitude and constraints on any available feature. On the left, two plans each one offering the possibility of accepting it or modifying it; on the right, one of the plans while being modified, all the actionable features can be accessed.
• Figure 3: Scores of the UEQ questionnaires for the two interfaces: the value 1 corresponds to the negative anchor and the value 5 corresponds to the positive anchor for the semantic differential items.