[Paper Review] Explaining Preference-driven Schedules: the EXPRES Framework

Alberto Pozanco, Francesca Mosca+3 more|arXiv (Cornell University)|Mar 16, 2022

Scheduling and Timetabling Solutions4 citations

TL;DR

This paper introduces EXPRES, a framework that explains why individual preferences remain unsatisfied in optimal, preference-driven schedules using Mixed-Integer Linear Programming (MILP) to identify the most relevant reasons. It translates these into human-readable explanations, demonstrating through simulations and user studies that employees at J.P. Morgan prefer automated explanations over human-generated ones.

ABSTRACT

Scheduling is the task of assigning a set of scarce resources distributed over time to a set of agents, who typically have preferences about the assignments they would like to get. Due to the constrained nature of these problems, satisfying all agents' preferences is often infeasible, which might lead to some agents not being happy with the resulting schedule. Providing explanations has been shown to increase satisfaction and trust in solutions produced by AI tools. However, it is particularly challenging to explain solutions that are influenced by and impact on multiple agents. In this paper we introduce the EXPRES framework, which can explain why a given preference was unsatisfied in a given optimal schedule. The EXPRES framework consists of: (i) an explanation generator that, based on a Mixed-Integer Linear Programming model, finds the best set of reasons that can explain an unsatisfied preference; and (ii) an explanation parser, which translates the generated explanations into human interpretable ones. Through simulations, we show that the explanation generator can efficiently scale to large instances. Finally, through a set of user studies within J.P. Morgan, we show that employees preferred the explanations generated by EXPRES over human-generated ones when considering workforce scheduling scenarios.

Motivation & Objective

• To address the challenge of explaining why certain preferences remain unsatisfied in optimal, multi-agent scheduling solutions, especially in constrained environments.
• To develop a scalable, automated explanation system that supports individual understanding while preserving agent privacy.
• To evaluate whether machine-generated explanations are more effective than human-generated ones in real-world scheduling contexts.
• To formalize the explanation generation task as an optimization problem that identifies the minimal set of influencing factors behind unsatisfied preferences.

Proposed method

• Model the explanation generation task as a Mixed-Integer Linear Programming (MILP) problem to find the minimal set of constraints and satisfied preferences that justify an unsatisfied preference.
• Define 'involved' and 'affected' agents and preferences to identify relevant factors influencing the unsatisfied preference.
• Use a two-stage process: first, generate low-level explanations via MILP; second, parse and group these into natural language explanations for end users.
• Preserve privacy by ensuring explanations only reveal necessary information about other agents’ preferences and assignments.
• Integrate the framework into a real-world workforce scheduling system at J.P. Morgan to evaluate usability and preference.
• Apply the framework to any scheduling problem with a totally ordered set of preference types, enabling generalization across domains.

Experimental results

Research questions

• RQ1 Can an automated system generate explanations for unsatisfied preferences in multi-agent scheduling that are both accurate and interpretable?
• RQ2 How does the scalability of the explanation generation component perform on large-scale scheduling instances?
• RQ3 Do users prefer machine-generated explanations over human-generated ones in real-world scheduling scenarios?
• RQ4 To what extent do explanations improve user understanding and trust in optimal schedules?
• RQ5 How can explanations be tailored to individual preferences and cognitive styles in a privacy-preserving manner?

Key findings

• The EXPRES framework successfully generates explanations for unsatisfied preferences using MILP, achieving efficient scaling to large scheduling instances.
• User studies at J.P. Morgan showed that employees preferred the automatically generated explanations from EXPRES over human-written ones in terms of clarity and perceived fairness.
• The explanation generator identified the most relevant constraints and satisfied preferences that caused a given preference to remain unsatisfied, providing actionable insight.
• The framework supports privacy by limiting the information revealed about other agents’ preferences and assignments.
• The explanation parser successfully translated low-level MILP outputs into natural language explanations that were both accurate and understandable to end users.
• The approach generalizes across different scheduling problems with a total order over preference types, demonstrating broad applicability beyond the initial use case.