What is it about?

Many automative desicion systems are implemented in networks, such as social networks, communication networks, financial transaction networks, or biological interaction networks, to make decisions. These systems can determine whether a loan/job application is approved, whether someone belongs to a particular group, or even which are the proper recommendations for an individual. However, these systems often fail to obtain the right decisions, and moreover to explain the rationale behind them or inform users about potential steps they can take to achieve a different outcome. This paper settle these problems by presenting a general framework for algorithmic recourse within networks. Instead of merely explaining why an individual-node possesses a certain property, our framework identifies the smallest realistic changes a user can make to alter that outcome. Importantly, the recommendations focus on actions that users can actually perform, such as modifying their own network connections, rather than suggesting arbitrary changes to the overall graph. We demonstrate this framework through the community detection case, aiming to provide suggestions to a node for leaving a specific community by making minimal adjustments to its connections, taking into account the maximum effort and willingness of an agent to perform changes. The framework is applicable to both weighted and unweighted networks, provides interpretable recommendations ranked by feasibility, and is supported by theoretical guarantees alongside extensive experimental validation. Furthermore, our method is evaluated against a Deep Reinforcement Learning Agent, which outperforms.

Featured Image

Why is it important?

As algorithms and automative models increasingly impact decisions in society, it is essential for people not only to understand these decisions but also to have meaningful ways to comperhend and overturn them. While existing explainable AI methods often address the question, "Why did this happen?" they fall short in answering the more practical question, "What can I do about it?" Our work expands the concept of algorithmic recourse to network-structured data, an area with very limited prior research. This is significant because many real-world systems, such as social media, recommendation systems, fraud detection, cybersecurity, and biological networks, function on graphs rather than traditional tabular data. By providing transparent and feasible recommendations, our framework helps restore user agency while avoiding opaque black-box methods. This methodology can support various applications, including privacy protection, reducing misinformation, improving fairness, analyzing network robustness, and understanding how local structural changes influence algorithmic decisions. More broadly, it shows that interpretable graph algorithms can achieve performance comparable to more complex learning-based methods while remaining easier to understand and trust.

Perspectives

We believe that future automative models and AI systems should not only explain their decisions but also empower people to take feasible actions based on those explanations. While algorithmic recourse has become a significant area of research in traditional machine learning, network structures pose unique challenges since individuals typically have control over only a small part of the graph. Our framework takes user agency into account, providing recommendations that are both practical and understandable. Although we showcase the framework through community detection, the core ideas are broadly applicable and can be adapted to various other graph-related problems, including influence analysis, centrality measures, fraud detection, recommendation systems, and graph neural networks. We hope this work inspires further research into actionable, human-centered explainable AI for networked data.

Chris Konstantopoulos
University of Patras

Read the Original

This page is a summary of: Algorithmic Recourse on Networks: From Explanations to Interventions, ACM Transactions on Knowledge Discovery from Data, July 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3831247.
You can read the full text:

Read

Contributors

The following have contributed to this page