RouteExplainer: An Explanation Framework for Vehicle Routing Problem

TL;DR

This paper tackles the lack of explainability in VRP solutions by introducing RouteExplainer, a post-hoc framework that treats a route as a sequence of edges and analyzes each edge's influence via an Edge Influence Model (EIM). It adds a many-to-many Transformer-based edge classifier to infer edge intentions, trains with a step-wise class-balanced loss, and uses LLMs to generate natural language explanations, including counterfactual routes for why-not questions. The framework is evaluated on multiple VRP datasets, showing rapid, accurate edge classification and producing qualitatively valid explanations, demonstrated through a tourist-route case study with GPT-4-based texts. The work demonstrates the practical potential of combining solver-agnostic explainability with LLM-powered narratives to support interactive, responsible VRP routing in real-world settings.

Abstract

The Vehicle Routing Problem (VRP) is a widely studied combinatorial optimization problem and has been applied to various practical problems. While the explainability for VRP is significant for improving the reliability and interactivity in practical VRP applications, it remains unexplored. In this paper, we propose RouteExplainer, a post-hoc explanation framework that explains the influence of each edge in a generated route. Our framework realizes this by rethinking a route as the sequence of actions and extending counterfactual explanations based on the action influence model to VRP. To enhance the explanation, we additionally propose an edge classifier that infers the intentions of each edge, a loss function to train the edge classifier, and explanation-text generation by Large Language Models (LLMs). We quantitatively evaluate our edge classifier on four different VRPs. The results demonstrate its rapid computation while maintaining reasonable accuracy, thereby highlighting its potential for deployment in practical applications. Moreover, on the subject of a tourist route, we qualitatively evaluate explanations generated by our framework. This evaluation not only validates our framework but also shows the synergy between explanation frameworks and LLMs. See https://ntt-dkiku.github.io/xai-vrp for our code, datasets, models, and demo.

Figures (12)

• Figure 1: Left: Action Influence Model causal_rl v.s. Edge Influence Model (ours). Right: the pipeline of RouteExplainer; it first takes a why and why-not question for VRP and simulates the CF route. The edge classifier then identifies the intentions of each edge in the actual and CF routes, and an LLM (e.g., GPT-4 openai2023gpt4) generates a counterfactual explanation by comparing the influences of the actual and CF edges.
• Figure 2: The proposed many-to-many edge classifier.
• Figure 3: The class ratio of edges w.r.t. steps, on each training split. The class ratios are for samples in which the number of visited nodes is the mode.
• Figure 4: Sequential confusion matrices that consist of $C\times C$ grids, where each grid visualizes the normalized confusion matrix value at each step as a heatmap in the order of steps from left to right. The top, middle, and bottom in each grid correspond to EC$_{\text{ce}}$, EC$_{\text{cbce}}$, and EC$_{\text{scbce}}$, respectively.
• Figure 5: