Holistic Explainable AI (H-XAI): Extending Transparency Beyond Developers in AI-Driven Decision Making

As AI systems increasingly mediate decisions in domains such as credit scoring and financial forecasting, their lack of transparency and bias raises critical concerns for fairness and public trust. Existing explainable AI (XAI) approaches largely serve developers, focusing on model justification rather than the needs of affected users or regulators. We introduce Holistic eXplainable AI (H-XAI), a framework that integrates causality-based rating methods with post-hoc explanation techniques to support transparent, stakeholder-aligned evaluation of AI systems deployed in online decision contexts. H-XAI treats explanation as an interactive, hypothesis-driven process, allowing users, auditors, and organizations to ask questions, test hypotheses, and compare model behavior against automatically generated random and biased baselines. By combining global and instance-level explanations, H-XAI helps communicate model bias and instability that shape everyday digital decisions. Through case studies in credit risk assessment and stock price prediction, we show how H-XAI extends explainability beyond developers toward responsible and inclusive AI practices that strengthen accountability in sociotechnical systems.

💡 Research Summary

The paper addresses a critical gap in the field of Explainable AI (XAI): most existing methods are designed for developers and model builders, leaving end‑users, regulators, and organizational stakeholders without adequate tools to understand AI‑driven decisions in high‑stakes domains such as credit scoring and financial forecasting. To bridge this gap, the authors introduce Holistic Explainable AI (H‑XAI), a framework that unifies traditional post‑hoc explanation techniques (e.g., SHAP, Partial Dependence Plots, counterfactuals) with a novel Rating‑Driven Explanation (RDE) component grounded in causal inference.

Stakeholder taxonomy and question mapping
Building on prior work, the authors categorize stakeholders into three groups: (1) individual users directly affected by model outputs, (2) regulatory bodies concerned with fairness and compliance, and (3) organizational actors (data scientists, business leads). They adopt an XAI question taxonomy that includes nine common query types (e.g., “Why was this prediction made?”, “What would change the outcome?”) and map each query to an appropriate explanation method: SHAP for feature attribution, counterfactuals for “what‑if” scenarios, and RDE for questions about bias, stability, or group‑level behavior.

Rating‑Driven Explanation (RDE)
RDE is the core novelty. It requires a causal graph that specifies relationships among a treatment variable (T) (e.g., loan amount requested), an outcome (O) (prediction or error), and a protected attribute (Z) (e.g., gender, age). The causal graph makes explicit any confounding pathways, allowing the framework to estimate de‑confounded treatment effects. Three quantitative metrics are defined:

1. Weighted Rejection Score (WRS) – aggregates pairwise t‑test rejections across demographic groups at multiple confidence levels (95 %, 75 %, 60 %) with decreasing weights, providing a concise measure of distributional disparity.
2. Average Treatment Effect (ATE) – estimates the expected change in (O) when intervening on (T) (e.g., halving the requested loan). Propensity‑score matching or G‑computation is used for estimation.
3. Deconfounded Impact Estimation (DIE %) – quantifies how much of the observed ATE is attributable to confounding by (Z). A large DIE % signals that apparent treatment effects are driven by protected attributes rather than the treatment itself.

The RDE workflow proceeds as follows: a user query is translated into a specification of (T), (O), and (Z); the appropriate metric is selected; scores are computed for the target model; and the results are compared against two automatically generated baselines—a random model (pure noise) and a biased model (predictions based solely on (Z)). This comparison grounds the explanation, making it easy for non‑technical stakeholders to interpret whether a model behaves more like a fair system, a random guesser, or an overtly discriminatory rule.

Case studies
Two empirical studies illustrate the framework.

Credit risk classification uses the German Credit dataset. RDE reveals that age and gender act as confounders influencing both the requested loan amount and the model’s approval decision. WRS highlights statistically significant disparities between age groups, while ATE shows that reducing the loan amount modestly improves approval odds for younger applicants but not for older ones. DIE % indicates that much of this effect is explained by age‑related confounding. Complementary SHAP visualizations pinpoint that “duration of existing credit” and “purpose of loan” are the strongest local drivers for individual decisions, satisfying the individual‑user query “Why was my loan denied?”.

Financial time‑series forecasting evaluates a stock‑price prediction model. Here, (O) is the absolute prediction error, and (T) is a synthetic market‑shock variable. RDE demonstrates that the model’s error distribution widens significantly for technology sector stocks after a shock (high ATE), and DIE % shows that this sensitivity is not merely due to sector‑specific volatility, confirming a genuine model instability. PDPs and SHAP explain which lagged features dominate predictions, addressing organizational questions about model performance.

Method selection guidance
Table 1 (referenced in the paper) provides a decision matrix linking stakeholder groups, question types, and recommended techniques. For example, regulators concerned with “Does the model systematically disadvantage a protected group?” are directed to RDE with WRS and DIE %; data scientists probing “How does a small input perturbation affect predictions?” should use ATE; end‑users asking “What would happen if I changed my loan amount?” are best served by counterfactuals combined with SHAP.

Contributions and implications
The authors claim three primary contributions: (1) a unified Holistic XAI framework that merges black‑box rating metrics with post‑hoc explanations, (2) a concrete RDE workflow that enables hypothesis testing, causal effect estimation, and baseline comparison, and (3) empirical evidence across classification and forecasting tasks that the combined approach satisfies diverse stakeholder needs. By moving explanation from a static, developer‑centric output to an interactive, hypothesis‑driven process, H‑XAI promises greater transparency, accountability, and trust in AI‑mediated decision making.

Future directions suggest extending H‑XAI to real‑time interactive dashboards, incorporating user‑studies to refine the usability of RDE visualizations, and applying the framework to other high‑impact domains such as healthcare triage or hiring algorithms. Overall, the paper makes a compelling case that holistic, stakeholder‑oriented explainability is essential for responsible AI deployment.