[Paper Review] Directive Explanations for Actionable Explainability in Machine Learning Applications

Ronal Singh, Paul Dourish+5 more|arXiv (Cornell University)|Feb 3, 2021

Explainable Artificial Intelligence (XAI)4 citations

TL;DR

This paper introduces directive explanations—actionable recommendations that tell individuals exactly what steps to take to achieve a desired outcome from a machine learning decision, such as loan approval. It presents an online study showing strong preference for directive explanations (p < 0.001) and proposes a Markov Decision Process-based conceptual model to generate such directives, emphasizing human-centered, context-specific recourse in AI systems.

ABSTRACT

This paper investigates the prospects of using directive explanations to assist people in achieving recourse of machine learning decisions. Directive explanations list which specific actions an individual needs to take to achieve their desired outcome. If a machine learning model makes a decision that is detrimental to an individual (e.g. denying a loan application), then it needs to both explain why it made that decision and also explain how the individual could obtain their desired outcome (if possible). At present, this is often done using counterfactual explanations, but such explanations generally do not tell individuals how to act. We assert that counterfactual explanations can be improved by explicitly providing people with actions they could use to achieve their desired goal. This paper makes two contributions. First, we present the results of an online study investigating people's perception of directive explanations. Second, we propose a conceptual model to generate such explanations. Our online study showed a significant preference for directive explanations ($p&lt;0.001$). However, the participants' preferred explanation type was affected by multiple factors, such as individual preferences, social factors, and the feasibility of the directives. Our findings highlight the need for a human-centred and context-specific approach for creating directive explanations.

Motivation & Objective

• To investigate whether directive explanations—explicit recommendations on actions to achieve a desired outcome—improve user understanding and recourse in machine learning decisions.
• To explore how individual preferences, social factors, and feasibility of actions influence the perception of directive versus non-directive explanations.
• To develop a conceptual model capable of generating directive explanations by framing recourse as a planning problem using Markov Decision Processes (MDPs).
• To evaluate the effectiveness of directive explanations in comparison to standard counterfactual explanations in real-world decision contexts such as loan approvals.
• To advocate for a human-centered, context-specific approach to explainable AI that prioritizes actionable, personalized guidance for users seeking recourse.

Proposed method

• Conducted an online study with 54 participants on Amazon MTurk, using 15 loan decision scenarios to compare three explanation types: non-directive counterfactuals, directive-specific (e.g., 'pay off your car loan'), and directive-generic (e.g., 'reduce your total debt').
• Collected ranked preferences and qualitative feedback on explanation types to assess user perception and reasoning.
• Proposed a conceptual model based on Markov Decision Processes (MDPs) to formalize the generation of directives as a planning problem from current state to desired outcome.
• Mapped actionable input features from trained models to real-world actions, ensuring recommendations are feasible and contextually relevant.
• Integrated user preferences and cost-of-action considerations into the explanation generation process, suggesting dialogue-based personalization for improved relevance.
• Used structural causal models and prior work on flip sets (Ustun et al., 2019) as foundational components to ensure generated directives lead to guaranteed desired outcomes.

Experimental results

Research questions

• RQ1 How do users perceive directive explanations compared to non-directive counterfactual explanations in terms of preference and perceived usefulness?
• RQ2 What factors—such as individual preferences, social context, or action feasibility—influence users’ choice between directive-specific and directive-generic explanations?
• RQ3 Can a conceptual model based on Markov Decision Processes effectively generate actionable, context-aware directives for recourse in machine learning decisions?
• RQ4 To what extent do users value explicit action recommendations over mere descriptions of counterfactual states in high-stakes decisions like loan approvals?
• RQ5 How can explanations be personalized to account for individual constraints and preferences while maintaining feasibility and correctness?

Key findings

• There was a statistically significant preference for directive explanations over non-directive counterfactuals (p < 0.001), with approximately 50% of participants selecting directive-specific explanations.
• Participants preferred directive-specific explanations when actions were feasible and aligned with personal circumstances, such as reducing debt or paying off a loan.
• Directive-generic explanations were favored when users wanted autonomy in choosing how to act, especially when specific actions were perceived as too costly or impractical.
• A significant number of participants expressed that explanations were less valuable when decisions were favorable (e.g., loan approved), highlighting context dependency in explanation utility.
• User preferences were influenced by perceived cost and feasibility of actions, indicating that not all directives are equally effective across individuals.
• The study revealed that a one-size-fits-all explanation type is insufficient, reinforcing the need for adaptive, personalized, and context-aware explanation generation in real-world AI systems.