🤖 machine learning

The Role of Causality in Algorithmic Recourse

This paper introduces a causal performative framework for algorithmic recourse that models how recommended actions propagate through a structural causal model to ensure genuine qualification improvements, thereby achieving stable equilibria that prevent strategic gaming and reduce the need for repeated model retraining.

Original authors: Srikanth Avasarala, Varun Gupta, Shahin Jabbari, Saber Salehkaleybar, Juba Ziani

Published 2026-07-31

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Srikanth Avasarala, Varun Gupta, Shahin Jabbari, Saber Salehkaleybar, Juba Ziani

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you're playing a video game where an invisible referee (an algorithm) decides if you win or lose based on your character's stats. Sometimes, the game tells you, "Hey, if you just add +5 to your strength, you'll pass the level!" This is the world of algorithmic recourse: giving people a clear, actionable checklist to improve their chances when a computer says "no" to things like getting a loan or a job. But here's the catch: what if the game is rigged? What if the "strength" stat you're told to boost is actually just a fake number that tricks the referee, while your real ability to survive the level stays the same? This is the problem of gaming the system.

Now, imagine that every time you change your stats, the referee actually changes the rules of the game to match your new look. This is called performative prediction. It's like a dance where the music changes every time you take a step, making it hard to find a stable rhythm. If the referee keeps changing the rules based on how you try to improve, the game becomes chaotic, and the advice you were given stops working. This paper dives into a specific corner of computer science where math meets human behavior, asking: How do we give advice that actually makes people better, rather than just tricking the computer, especially when the computer is watching and reacting to our moves?

The Trap of the "Fake" Fix

In the world of high-stakes decisions—like banks deciding who gets a mortgage or companies deciding who gets hired—machines often make the call. When a machine rejects an application, it's helpful to get a "recourse" plan: a list of changes to your life that would flip the result from "No" to "Yes." For example, "Pay off $500 of debt, and you'll get approved."

But there's a sneaky problem. Most current methods just look at the math of the machine. They say, "If you change feature X, the machine's score goes up." They don't care if that change actually makes you a better borrower or a more qualified employee. You might be "gaming" the system. Imagine you open ten fake credit card accounts just to boost your credit score number. The machine sees the higher number and says "Yes!" But you haven't actually become more financially responsible; you've just tricked the calculator.

When people start doing this, the machine gets confused. It sees a bunch of people with high scores who can't pay their bills. The machine's predictions become useless. Then, the bank has to retrain the machine with new rules. But now, the old advice ("open fake cards") doesn't work anymore. The goalposts have moved, and the advice is broken. This is a vicious cycle where the solution makes the problem worse.

The Causal Compass

This paper introduces a new way to think about this problem using causality. Instead of just looking at the numbers, the authors ask: "What actually causes the result?"

Think of your features (like income, debt, education) as a network of pipes. If you pour water into one pipe (an intervention), it flows through the network and changes the water level in other pipes.

The Old Way: "If you increase the water in Pipe A, the machine thinks the tank is full." It ignores that Pipe A might be leaking into Pipe B, which is actually the one that matters.
The New Way (Causal): "If you pour water into Pipe A, it flows through the pipes and actually fills the tank."

The authors build a model where they map out these pipes (a "Structural Causal Model"). They understand that changing one thing (like getting a job) might naturally lead to another change (like having more savings). They use this map to figure out which actions will genuinely improve a person's true qualifications, not just their score.

The Dance of the Algorithm

Here is where it gets tricky. The paper treats the whole situation as a dance between the Learner (the bank building the model) and the Agents (the people applying for loans).

The Learner sets a rule.
The Agents try to change their features to pass the rule.
Because of the causal links, their changes shift the whole data landscape.
The Learner sees this new landscape and has to update the rule.

The authors show that if you ignore the causal links, this dance becomes a mess. The math becomes "non-convex," which is a fancy way of saying the path to the solution is full of bumps and dead ends, and standard math tools can't find the way out. It's like trying to find the bottom of a valley that keeps moving and reshaping itself.

However, the paper proves that if you account for the causal structure, you can find a stable solution. This is a point where the Learner's rule and the Agents' actions settle into a rhythm. The Agents make real improvements, the Learner's model stays accurate, and nobody is tricking the system.

What They Found

The researchers tested their ideas using two types of data:

A made-up world: They created a fake loan dataset with 7 different features (like age, education, income) and programmed specific rules for how they affect each other.
Real-world data: They used a real dataset of 30,000 credit card clients from Taiwan.

They used two main strategies to find the stable solution:

Repeated Risk Minimization (RRM): Like a teacher who gives a test, grades it, adjusts the lesson plan, gives the test again, and repeats until the class and the teacher agree.
Repeated Gradient Descent (RGD): A slightly faster, step-by-step version of the same process.

The Results:

Convergence: In their simulations, these methods successfully found a stable point where the model stopped changing wildly. The math showed that as long as the "gaming" wasn't too strong (controlled by a parameter called $\kappa$ ), the system would settle down.
Beating the Naive Approach: When they compared their causal method to the "naive" method (which ignores causality and just looks at the score), the causal method was much better. As the agents tried harder to game the system, the naive method got worse and worse, while the causal method kept the model accurate.
The Cost of Change: They also looked at how expensive it is to change different features. If it's cheap to change "upstream" features (like getting a degree), the agents do that. If it's cheap to change "downstream" features (like saving money), they do that. The causal model correctly predicted how these costs would ripple through the system to change the final outcome.

The Takeaway

The paper doesn't claim to have solved every problem in the world. It shows, through math and computer simulations, that ignoring how things actually cause each other leads to broken advice. When we give people a checklist to improve their lives, we need to make sure that checklist actually improves their lives, not just their score on a screen.

By using a "causal map," we can design systems where the advice given to a person leads to real, lasting improvement. This stops people from trying to trick the system and keeps the machine's predictions honest. It turns a game of "gotcha" into a partnership where everyone wins: the person gets a fair shot, and the bank gets a reliable customer. The authors suggest that this approach is a necessary step to stop the "moving goalposts" problem and make AI decisions that are both fair and effective.

Technical Summary: The Role of Causality in Algorithmic Recourse

Problem Statement

Algorithmic recourse aims to provide individuals with actionable recommendations to improve their predicted outcomes in high-stakes settings (e.g., loan approvals). However, existing approaches typically optimize solely for changing the deployed model's prediction, ignoring whether the recommended actions genuinely improve an agent's true qualifications or merely "game" the system. This misalignment can lead to strategic behavior where agents manipulate non-causal features to obtain better scores without improving their underlying ability (e.g., opening dummy credit cards). Consequently, deployed models become less accurate over time, necessitating repeated retraining and rendering recourse recommendations ineffective as "moving goalposts."

This paper addresses the failure mode where recourse policies incentivize gaming by explicitly incorporating causality into the recourse framework. The core challenge is to jointly design a predictive model and a recourse policy that promote meaningful improvement while maintaining predictive accuracy, accounting for the fact that recommended actions shift the data distribution.

Methodology

Causal Performative Framework

The authors formalize the interaction between a learner and strategic agents using a Structural Causal Model (SCM).

Feature Space: Agents possess features $x \in \mathbb{R}^d$ and true labels $y$ .
Causal Structure: Dependencies are modeled via a Directed Acyclic Graph (DAG) $G=(V, E, \omega)$ . Interventions on one feature propagate through the graph, affecting downstream features and the true label.
Intervention Propagation: An intervention vector $a$ induces feature changes $\Delta x(a) = G_x^\top a$ and label changes $\Delta y(a) = G_y^\top a$ , where $G_x$ and $G_y$ are contribution matrices derived from the causal graph.
Agent Response: Agents choose a recourse action $a^*$ to maximize utility: $w^\top(x + \Delta x(a)) - \frac{1}{\kappa}c(a)$ , where $c(a)$ is the intervention cost and $\kappa$ controls intervention strength. Under quadratic costs, this yields a closed-form solution $a^* = \frac{\kappa}{2} C^{-1} G_x w$ .

Performative Prediction Formulation

The learning problem is framed as performative prediction, where the model $w$ influences the data distribution $D_w$ through agent responses.

Objective: The learner seeks to minimize the expected loss under the induced distribution: $\min_w \mathbb{E}_{(x,y) \sim D_w} [\ell(w^\top x, y)]$ .
Non-Convexity: Even with convex loss functions and quadratic costs, the resulting optimization problem is generally non-convex due to the feedback loop between the model and the induced distribution shift.
Stability vs. Optimality: The paper focuses on finding performatively stable solutions ( $w_s$ ), defined as fixed points where retraining on the distribution induced by $w_s$ yields $w_s$ again. This is distinct from, but close to, the performatively optimal solution ( $w^*$ ).

Algorithms

Two iterative algorithms are proposed to compute stable solutions:

Repeated Risk Minimization (RRM): At each iteration, the learner collects a sample, applies the causal response model to generate an induced dataset, and re-trains the model to minimize empirical risk on this new dataset.
Repeated Gradient Descent (RGD): A gradient-based approximation where the learner performs a single gradient descent step on the empirical objective of the induced dataset at each iteration.

Theoretical Guarantees

The authors establish conditions for the convergence of RRM and RGD to a stable solution. Convergence depends on the performative sensitivity, governed by the parameter $\kappa \Lambda$ , where $\Lambda$ depends on the causal propagation matrices ( $G_x, G_y$ ) and the cost matrix ( $C$ ).

If the intervention strength $\kappa$ is sufficiently small relative to the strong convexity and smoothness constants of the loss function, the algorithms converge to a stable point.
The stable solution is shown to be near-optimal (close to $w^*$ ) under specific conditions.

Key Contributions

Causal Performative Framework: Introduction of a framework that models recourse interventions as propagating through a structural causal model, capturing both direct and indirect effects on features and true labels.
Formulation as Performative Prediction: Characterization of the causal recourse problem as a performative prediction problem, demonstrating that standard convex optimization techniques fail due to non-convexity induced by strategic responses.
Algorithmic Solutions: Development of RRM and RGD algorithms with formal convergence guarantees for computing performatively stable recourse policies.
Empirical Validation: Demonstration across semi-synthetic and real-world credit datasets that accounting for causal responses reduces incentives for gaming and improves predictive performance compared to naive approaches.

Experimental Results

The authors evaluated the framework on:

A 7-dimensional semi-synthetic loan dataset generated from a known SCM.
The Taiwan Credit Card dataset, aggregated into 6 features with a causal graph estimated via regression.

Key Findings:

Convergence: Both RRM and RGD converge to stable solutions. RRM converges rapidly for small $\kappa$ but may fail to converge for large $\kappa$ (consistent with theoretical bounds), while RGD exhibits consistent monotonic decay.
Performative Improvement: As the performativity parameter $\kappa$ increases (stronger strategic adaptation), the performance gap between the naive Expected Risk Minimization (ERM) baseline and the proposed stable/optimal models widens. The proposed methods significantly outperform naive ERM in terms of the performative objective.
Causal Propagation: The experiments show that intervention costs dictate where agents act, but the causal graph dictates how those actions propagate. Even when downstream features have infinite intervention costs (immutable), they still change significantly due to upstream interventions propagating through the causal structure. This highlights the necessity of modeling $G_x$ rather than treating features as independent.
Robustness: Results remain consistent across different intervention cost profiles (Uniform, Upstream-cheap, Downstream-cheap).

Significance and Claims

The paper claims that ignoring causal structure in algorithmic recourse leads to large, misaligned behavioral responses that degrade model accuracy and render recourse ineffective. By contrast, causal recourse leads to equilibria that lower the impact of gaming.

The authors emphasize that their work bridges the gap between strategic classification (designing robust models) and algorithmic recourse (providing actionable advice). They argue that for recourse to be truly effective and sustainable, it must account for how interventions shift the underlying data distribution and true labels, not just the model's prediction. The proposed framework provides a mathematically grounded method to compute stable policies that incentivize genuine improvement rather than manipulation.

The paper modestly notes that it assumes access to a known structural causal model and suggests future work should address learning the causal graph jointly with the recourse policy or handling uncertainty in the graph.