Learning-Augmented Robust Algorithmic Recourse

TL;DR

The paper investigates recourse when predictive models evolve over time and introduces a learning-augmented framework that uses predictions $\hat{\theta}$ of the future model to balance robustness against model changes with low recourse cost. It formulates robustness $R(x',\alpha)$ and consistency $C(x',\hat{\theta})$, and optimizes $\min_{x'} \beta R(x',\alpha) + (1-\beta) C(x',\hat{\theta})$ to navigate the trade-off, presenting Algorithm \ref{alg:l1-linf} that relies on a local linear approximation (LIME) and is provably optimal for generalized linear models when $\beta \in \{0,1\}$. Empirically, the method Pareto-dominates ROAR in robustness and consistency across synthetic and real datasets, with results showing how the quality of the future-model prediction $\hat{\theta}$ modulates smoothness and validity-cost trade-offs. The work advances practical, prediction-aware recourse in dynamic settings and motivates future work on actionability, alternative model-change notions, and stronger theoretical guarantees.

Abstract

Algorithmic recourse provides individuals who receive undesirable outcomes from machine learning systems with minimum-cost improvements to achieve a desirable outcome. However, machine learning models often get updated, so the recourse may not lead to the desired outcome. The robust recourse framework chooses recourses that are less sensitive to adversarial model changes, but this comes at a higher cost. To address this, we initiate the study of learning-augmented algorithmic recourse and evaluate the extent to which a designer equipped with a prediction of the future model can reduce the cost of recourse when the prediction is accurate (consistency) while also limiting the cost even when the prediction is inaccurate (robustness). We propose a novel algorithm, study the robustness-consistency trade-off, and analyze how prediction accuracy affects performance.

Figures (11)

• Figure 1: The trade-off between robustness and consistency for $\alpha=0.5$: rows and columns correspond to different datasets and models as indicated in the sub-caption. In each subfigure, each curve shows the trade-off for different predictions for our algorithm and ROAR.
• Figure 2: The smoothness for predictions with different accuracies: rows and columns correspond to different datasets and models as indicated in the sub-caption. In each subfigure, curves correspond to different predictions and track the smoothness as a function of $\beta$ for the given prediction.
• Figure 3: The trade-off between future validity and cost: rows and columns correspond to different datasets and models. In each subfigure, curves show the trade-off for different algorithms.
• Figure 4: The trade-off between robustness and consistency for $\alpha=0.5$ with error bars for robustness: logistic regression (left) and neural network (right). Rows correspond to datasets: synthetic (top), German (middle), and Small Business (bottom). In each subfigure, each curve shows the trade-off for different predictions. The robustness and consistency of ROAR solutions at $\beta=1$ are mentioned in parentheses and depicted by stars. Missing stars are outside the range of the figure.
• Figure 5: The trade-off between robustness and consistency for $\alpha=0.5$ with error bars for consistency: logistic regression (left) and neural network (right). Rows correspond to datasets: synthetic (top), German (middle), and Small Business (bottom). In each subfigure, each curve shows the trade-off for different predictions. The robustness and consistency of ROAR solutions at $\beta=1$ are mentioned in parentheses and depicted by stars. Missing stars are outside the range of the figure.

Theorems & Definitions (10)

• Definition 3.1: Robustness
• Definition 3.2: Consistency
• Theorem 3.3
• Lemma C.2
• proof
• Lemma C.3
• proof
• Lemma C.4
• proof
• proof : Proof of Theorem \ref{['thm:opt-l1-linf']}