Personalized Algorithmic Recourse with Preference Elicitation

TL;DR

PEAR tackles the problem of personalized algorithmic recourse by learning user-specific action costs through Bayesian preference elicitation and integrating this with AR via reinforcement learning and Monte Carlo Tree Search. It introduces a cost correlation structure to capture interactions between actions and uses choice-set queries optimized for information gain (Expected Utility of Selection) to rapidly refine user weights ${\bm{w}}$. A key contribution is the W-FARE framework, which personalizes recourse by marginalizing costs over the posterior to guide efficient, user-tailored interventions, and its integration with MCTS accelerates convergence. Empirical results on real-world datasets show substantial gains in validity and cost savings over non-personalized baselines, with robustness to partial knowledge about cost structures and extensions like XPEAR providing explainable, user-aware interventions. The work advances practical, interactive recourse with real-world impact for high-stakes decisions while highlighting ethical considerations around fairness and privacy.

Abstract

Algorithmic Recourse (AR) is the problem of computing a sequence of actions that -- once performed by a user -- overturns an undesirable machine decision. It is paramount that the sequence of actions does not require too much effort for users to implement. Yet, most approaches to AR assume that actions cost the same for all users, and thus may recommend unfairly expensive recourse plans to certain users. Prompted by this observation, we introduce PEAR, the first human-in-the-loop approach capable of providing personalized algorithmic recourse tailored to the needs of any end-user. PEAR builds on insights from Bayesian Preference Elicitation to iteratively refine an estimate of the costs of actions by asking choice set queries to the target user. The queries themselves are computed by maximizing the Expected Utility of Selection, a principled measure of information gain accounting for uncertainty on both the cost estimate and the user's responses. PEAR integrates elicitation into a Reinforcement Learning agent coupled with Monte Carlo Tree Search to quickly identify promising recourse plans. Our empirical evaluation on real-world datasets highlights how PEAR produces high-quality personalized recourse in only a handful of iterations.

Figures (7)

• Figure 1: Overview of PEAR. (1) Given the initial state ${\bm{\mathrm{s}}}^{(0)}$ of the user and weights ${\bm{\mathrm{w}}}^{(0)}$, PEAR computes a pool of candidate interventions achieving recourse. (2) A choice set $O^{(t)}$ is selected from the pool and presented to the user. (3) The user picks their preferred intervention from the set. (4) An improved estimate of the weights ${\bm{\mathrm{w}}}^{(t+1)}$ is computed using this feedback, and (5) the user's state ${\bm{\mathrm{s}}}^{(t+1)}$ is updated. After $T$ rounds, the estimated weights are used to compute a final intervention $I^{*}$.
• Figure 2: The cost model (Left) A cost correlation structure (CCS) for cost modelling. (Right) Given ${\bm{\mathrm{s}}}^{(0)} = [1,1,1]$ and unit ${\bm{\mathrm{w}}}$, let us imagine we want to reach ${\bm{\mathrm{s}}}^{(3)} = [5,5,5]$ by following the presented CCS and the intervention $I=\{a_1, a_2, a_3\}$, where $a_i$ assign $s_i \leftarrow 5$ for all $i \in \{1,2,3\}$. Clearly, we incur different costs by applying permuted versions of $I$. The green path indicates the lower-cost intervention.
• Figure 3: Normalized Average Regret for PEAR when varying the number of questions, the choice set size and the user response model on both datasets (sampled from All users).
• Figure 4: Directed Acyclic Graphs used for the Cost Correlation Structures. Each node corresponds to a specific user feature. The complete mapping between node acronyms and features can be found in the experimental configuration files.
• Figure 5: The W-EFARE method. Given $({\bm{\mathrm{s}}}', {\bm{\mathrm{w}}})$, we traverse the graph (green path) using the decision tree (orange components) to choose the next action $(f,x)^{(t)}$ from the available transitions. When we reach the STOP node, we return the found intervention $I^*$. On the right, we have a generated intervention achieving recourse for the GiveMeSomeCredit dataset. On the top-right, we have the Boolean rule extracted from the decision tree motivating the first recommended action $\mathtt{ReduceOpenCreditLines}$. As we can see, W-EFARE can generate rules that also depend on the cost function $C$, achieving the desired user awareness.

Theorems & Definitions (1)

• Definition 1: Personalized Algorithmic Recourse