TL;DR
This paper investigates the temporal robustness of algorithmic recourse, revealing that existing methods often fail over time unless the environment is stationary, and proposes a new temporal approach to improve resilience.
Contribution
It introduces a causal framework for analyzing recourse over time and proposes a novel algorithm that explicitly incorporates temporal dynamics for more durable solutions.
Findings
Causal AR methods can fail over time in non-stationary environments.
Counterfactual AR cannot be optimally solved unless the world is deterministic.
The proposed temporal AR algorithm improves resilience in synthetic and real datasets.
Abstract
Algorithmic Recourse (AR) aims to provide users with actionable steps to overturn unfavourable decisions made by machine learning predictors. However, these actions often take time to implement (e.g., getting a degree can take years), and their effects may vary as the world evolves. Thus, it is natural to ask for recourse that remains valid in a dynamic environment. In this paper, we study the robustness of algorithmic recourse over time by casting the problem through the lens of causality. We demonstrate theoretically and empirically that (even robust) causal AR methods can fail over time, except in the -- unlikely -- case that the world is stationary. Even more critically, unless the world is fully deterministic, counterfactual AR cannot be solved optimally. To account for this, we propose a simple yet effective algorithm for temporal AR that explicitly accounts for time under the…
Peer Reviews
Decision·Submitted to ICLR 2025
1. The paper studies an interesting phenomenon: the temporal dimension of algorithmic recourse. It is a relevant problem since implementing recourses can take a long time. Thus, we need to have a mechanism to take the time-horizon (in this paper is $\tau$) into the recourse problem.
The main downsides of this approach are 1. It requires an estimator $\tilde P(X^t)$. I would argue that in most practical settings, this distribution is never available. If it is available, it could be easily misspecified as well. I believe there is a big gap between what the authors think are practical, and what I think are practical. It is extremely unclear to me how I could use the proposed method in a real-world deployment. The problem is complex when time is taken into account, and having
The concept of Temporal Sub-population Algorithmic Recourse (T-SAR) adds a unique dimension by incorporating time-awareness, which is relatively novel in the field of causal AR. The theoretical contributions are well-grounded in causal inference, and the authors provide a robust theoretical framework that explains why existing AR models may fail over time. By addressing the challenges of AR in non-static environments, this work highlights an important area for improvement in AR techniques. The e
The explanation of how temporal sensitivity interacts with causality is unclear. For instance, while the paper mentions the impact of time on causal interventions, it doesn’t sufficiently explain how different timing of interventions might lead to varied results, especially in non-stationary environments. This leaves readers uncertain about the long-term implications of time on AR effectiveness, as the dynamic causal effects of time-sensitive actions are only superficially addressed. The paper
This paper investigate an important but not studied problem, that is the influence of time on algorithmic recourse. This paper explore the problem from the pespective of both theory and empirical examinations. Extensive theoretical analysis is sound and comprehensive. The empirical improvement in experiements is significant, showing the effectiveness of the proposed method.
1. The presentation of the paper could be further improved. The problem formulation is not clear and kind of vague. I guess the problem is to determine the intervention $\theta$ at time $t$ and implement it at a later time $t+\tau$ instead of determining intervention at $t+\tau$. However, it is not distinguished in the manuscript. And the notation $do(\theta)$ is confusing, because it is not clear that at which time step the intervention $\theta$ is conducted. 2. I think in the field of algorith
Code & Models
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
