Zoomable Level-of-Detail ChartTables for Interpreting Probabilistic Model Outputs for Reactionary Train Delays

TL;DR

The paper addresses interpreting probabilistic outputs from a stochastic Monte-Carlo-style agent-based model of reactionary delay in UK railways. It introduces Zoomable Level-of-Detail ChartTables that place composite metrics in a case-by-variable table, with mini-charts providing two levels of detail and interactive features such as semantic zoom, filtering, sorting, and highlighting. The main contributions include four mini-chart types that capture different aspects of delay distributions, a design that supports distribution-aware interpretation, and discussion of industry relevance and potential broader uses. The approach supports timetable robustness testing and helps identify problematic trains and underlying causes, with positive feedback from train-operating companies and licensing to RSSB. The work proposes generalizable visual-analytics strategies for exploring probabilistic, multivariate outputs in large-scale systems.

Abstract

"Reactionary delay" is a result of the accumulated cascading effects of knock-on train delays which is increasing on UK railways due to increasing utilisation of the railway infrastructure. The chaotic nature of its effects on train lateness is notoriously hard to predict. We use a stochastic Monte-Carto-style simulation of reactionary delay that produces whole distributions of likely reactionary delay and delays this causes. We demonstrate how Zoomable Level-of-Detail ChartTables - case-by-variable tables where cases are rows, variables are columns, variables are complex composite metrics that incorporate distributions, and cells contain mini-charts that depict these as different levels of detail through zoom interaction - help interpret whole distributions of model outputs to help understand the causes and effects of reactionary delay, how they inform timetable robustness testing, and how they could be used in other contexts.

Figures (5)

• Figure 1: Four metric types depicted by four different mini-charts. Each has both low and high level-of-detail (LoD) variants. In all cases here, the high LoD variant on the right shows the detail of the top five rows of the low LoD on the left (outlined in a blue dotted line). An appropriate threshold then vertically zooming rows determines which variant is used.
• Figure 2: Excerpt zoomed out to the top 20% of delay-causing trains, sorted by delay caused showing the "drop-off". Only a few trains are suffering significant reactionary delay. This is caused by only a few other trains, but they are not causing much reactionary delay to other services and are not themselves delayed much nor causing passenger journey disruption. Trains suffering 10-20% of the total reactionary delay go on to cause delays to other services.
• Figure 3: Four examples of high LoD "LatenessTProfileCharts" (Fig. \ref{['fig:types']}b) showing different lateness consistency between model runs.
• Figure 4: Example sorted by "average station stop lateness" with observations made on other aspects of lateness for these trains.
• Figure 5: "Highlighting" interaction that identifies two of the trains involved with the train identified with the mouse cursor.