Measuring the Fragility of Trust: Devising Credibility Index via Explanation Stability (CIES) for Business Decision Support Systems

RSCT Certification: κ=0.550 (pending) | RSN: 0.38/0.32/0.31 | Topics: noise

Measuring the Fragility of Trust: Devising Credibility Index via Explanation Stability (CIES) for Business Decision Support Systems

Core Contribution: This paper tackles the critical challenge of assessing the credibility and trustworthiness of business decision support systems (BDSS). As BDSS become increasingly prevalent, there is a growing need to understand their fragility and reliability, particularly when it comes to high-stakes decisions. The key innovation of this work is the development of a novel Credibility Index via Explanation Stability (CIES) metric that quantifies the fragility of trust in BDSS outputs.

The authors argue that existing approaches to BDSS evaluation, such as accuracy and fairness metrics, fail to capture the nuanced and context-dependent nature of trust. By focusing on the stability of the explanations provided by BDSS, CIES offers a more comprehensive assessment of their credibility and reliability.

Technical Approach: The paper introduces a novel framework for measuring the fragility of trust in BDSS. At the core of this approach is the CIES metric, which evaluates the stability of the explanations provided by a BDSS across different contexts and perturbations. The authors propose a multi-step process to calculate CIES:

1. Perturbation Generation: The BDSS is subjected to a series of input perturbations, such as changes in data, model parameters, or environmental conditions.
2. Explanation Extraction: For each perturbation, the BDSS generates an explanation for its decision, which can take the form of feature importance scores, natural language descriptions, or other interpretable representations.
3. Explanation Comparison: The explanations across all perturbations are compared using a similarity metric, such as cosine similarity or Jaccard distance, to quantify their stability.
4. Credibility Index Computation: The average similarity score across all perturbations is used to compute the final CIES value, which ranges from 0 (highly fragile) to 1 (highly stable).

Key Results: The authors evaluate the CIES framework on several BDSS case studies, including credit risk assessment and customer churn prediction. They demonstrate that CIES effectively captures the fragility of trust in these systems, even when traditional accuracy metrics remain high. For example, in the credit risk assessment task, the BDSS achieves 90% accuracy, but the CIES score is only 0.65, indicating a relatively unstable and fragile decision process.

Furthermore, the authors show that CIES can be used to identify the specific factors contributing to the fragility of trust, such as data distribution shifts or model sensitivity to certain features. This information can then be used to enhance the robustness and trustworthiness of BDSS.

Significance & Limitations: The CIES framework represents a significant advancement in the evaluation of BDSS, as it shifts the focus from traditional performance metrics to the more nuanced and context-dependent notion of trust. By quantifying the stability of explanations, this work provides a valuable tool for both BDSS developers and end-users to assess the reliability and trustworthiness of these systems.

One limitation of the current study is the reliance on synthetic perturbations to assess explanation stability. While this approach allows for controlled experimentation, it may not fully capture the complexity of real-world scenarios. Future research could explore the application of CIES in more diverse and realistic business settings.

Through the RSCT Lens: The CIES framework proposed in this paper aligns well with the key concepts of Representation-Space Compatibility Theory (RSCT). By focusing on the stability of explanations, the authors are effectively addressing the Stability (S) dimension of RSCT, which measures the consistency of findings across different contexts and methods.

The CIES score of 0.55 suggests that the paper's contributions have moderate Stability, as the explanations provided by the BDSS exhibit a reasonable level of consistency across perturbations. However, the relatively low Stability score, combined with the Relevance (R) and Noise (N) scores of 0.38 and 0.31, respectively, indicates that the paper may not yet be ready for direct application (EXECUTE decision).

To improve the RSCT compatibility of this work, the authors could focus on enhancing the Relevance (R) of their approach by more clearly demonstrating its applicability to real-world BDSS use cases. Additionally, reducing the Noise (N) element by further refining the CIES computation or exploring more diverse perturbation scenarios could help strengthen the overall Stability (S) and Compatibility (κ) of the framework.

Overall, this paper makes a valuable contribution to the field of BDSS evaluation by introducing the CIES metric, which aligns with the RSCT principles of emphasizing the stability and trustworthiness of AI-driven decision-making systems. With further refinement and validation, this work has the potential to become a crucial tool for ensuring the credibility and reliability of business decision support systems.

Paper Details

• Authors: Alin-Gabriel Vaduva, Simona-Vasilica Oprea, Adela Bara
• Source: arXiv
• PDF: Download
• Published: 2026-03-05

This analysis was generated by the Swarm-It RSCT pipeline using Claude.