TL;DR
This paper introduces a novel model for accountable ordinal sorting in multi-criteria decision-making. It addresses the need for transparent and trustworthy recommendations by designing a system that provides convincing explanations, allowing users to defend the system's output as their own.
What Problem Does This Paper Address?
In today's data-driven world, automated decisions and recommendations are ubiquitous. However, there's a growing demand for these systems to be transparent and trustworthy, especially with regulations like GDPR emphasizing a "right to explanation." While existing approaches focus on trust (accuracy) or transparency (access to algorithms), this paper tackles "accountability." Accountability goes beyond trust and transparency; it means the recipient of a recommendation can genuinely defend it to skeptical stakeholders. This requires a deeper understanding of why a recommendation is made, not just what it is. The paper specifically addresses ordinal sorting problems, where objects are ranked into ordered categories, and the challenge of providing explanations that are both faithful to the underlying model and understandable to the user.
Methods
The authors propose a recommender system designed with two core principles: "no jargon" and "no arbitrariness."
- No Jargon: Instead of asking users about abstract model parameters (like weights), the system elicits preferences directly through examples. Users assign typical objects to categories, and the system learns from these assignments.
- No Arbitrariness: The system avoids picking a single "representative" parameter value. Instead, it formulates recommendations that are robust and cannot be refuted by any consistent interpretation of the user's preferences.
Additionally, the model incorporates three assumptions to enhance human reasoning:
- No Compensation: Criteria evaluations are treated ordinally; good performance on one criterion cannot compensate for poor performance on another.
- No Values: The system uses outranking relations rather than value functions, avoiding the need to assign numerical values to preferences.
- No Frontiers: The model avoids explicit "frontiers" between classes, instead using simple classification rules that compare a new object to already classified reference objects.
The system's output includes not only the recommended class(es) but also a formal explanation, structured as argument schemes, to support the recommendation.
Datasets
Not discussed in the paper. The paper uses synthetic examples to illustrate its concepts.
Key Insights and Ideas
- Accountability as the Goal: The paper clearly distinguishes accountability from trust and transparency, defining it as the ability to defend a recommendation. This reframes the goal of explainable AI.
- Indirect Elicitation is Key: Learning preferences through direct examples rather than abstract parameters makes the system more intuitive and less prone to user confusion.
- Robustness through Non-Arbitrariness: By avoiding single-point estimations of preferences, the system provides recommendations that are more resilient to incomplete or slightly inconsistent user input.
- Argumentation for Explanations: Structuring explanations as formal argument schemes (inspired by Walton's work) provides a principled way to justify recommendations, making them understandable and defensible.
- "Everything is possible, unless proven otherwise": This principle guides the explanation generation, focusing on why certain assignments are impossible rather than trying to prove why others are possible.
Key Results
The paper presents a formal model for accountable ordinal sorting and demonstrates how it can generate explanations. The key results are the design principles and the framework for generating explanations based on argument schemes. The paper provides examples illustrating how the model handles different scenarios, including identifying possible and impossible assignments for candidate objects based on preference information. For instance, in Example 5, it shows how the system can explain why a candidate "X" cannot be ranked higher than a certain class, citing the comparison with a reference object "B1."
Limitations and Future Work
- Handling Inconsistencies: The authors acknowledge that their model, while simple, can lead to inconsistencies. They suggest two avenues for future work: relaxing some of the design assumptions (e.g., allowing frontiers) or handling inconsistencies using defeasible and non-monotonic reasoning techniques.
- Formal Argumentation: The paper points to formal argumentation frameworks as a promising direction for handling conflicting arguments and determining acceptable ones, potentially leading to more robust decision-making.
- Dialogue Games: Exploring argumentation systems as dialogue games could keep the decision-maker actively involved and allow for more nuanced explanations.
- Analogical Reasoning: The paper notes that some of the argument schemes have a flavor of analogical reasoning, suggesting further research in this area for explanation generation.
Summary updated