ORDENA: ORigin-DEstiNAtion data exploration

TL;DR

The paper addresses the challenge of analyzing large-scale origin-destination OD flows where aggregation can obscure patterns. It introduces ORDENA, a visual analytics system built around the OD-Plot, which orders origins and destinations via the Fiedler vector to reveal spatial patterns without data aggregation and couples this with a classification/explanation pipeline using SHAP for anomaly assessment. The work demonstrates ORDENA through three case studies on real-world Melbourne and São Paulo datasets, validated by expert evaluations, highlighting its ability to uncover global/local patterns, explain influential features, and assess region-specific model performance. The proposed framework integrates visualization, interpretable machine learning, and explainability in a unified platform, offering practical tooling for transportation analytics, crime mapping, and logistics risk assessment.

Abstract

Analyzing origin-destination flows is an important problem that has been extensively investigated in several scientific fields, particularly by the visualization community. The problem becomes especially challenging when involving massive data, demanding mechanisms such as data aggregation and interactive filtering to make the exploratory process doable. However, data aggregation tends to smooth out certain patterns, and deciding which data should be filtered is not straightforward. In this work, we propose ORDENA, a visual analytic tool to explore origin and destination data. ORDENA is built upon a simple and intuitive scatter plot where the horizontal and vertical axes correspond to origins and destinations. Therefore, each origin-destination flow is represented as a point in the scatter plot. How the points are organized in the plot layout reveals important spatial phenomena present in the data. Moreover, ORDENA provides explainability resources that allow users to better understand the relation between origin-destination flows and associated attributes. We illustrate ORDENA's effectiveness in a set of case studies, which have also been elaborated in collaboration with domain experts. The proposed tool has also been evaluated by domain experts not involved in its development, which provided quite positive feedback about ORDENA.

Figures (12)

• Figure 1: Visualization of spatial units characterized by the Fiedler vector (zip codes or street intersections). On the left, zip code tracts from the Melbourne accidents dataset; on the right, street intersections from the São Paulo bicycle theft dataset. In both maps, units are color-coded according to their Fiedler vector indices, with nearby units showing similar colors and the chromatic scales displayed using a continuous colormap.
• Figure 2: OD-Plot interpretation and expressiveness. The arrangement of points corresponding to OD flows on the plot reveals different spatial patterns.
• Figure 3: OD-plot (left) and matrix representations with resolutions 18$\times$18 (middle) and 71$\times$71 (right). OD patterns are more clearly revealed on the OD-plot.
• Figure 4: Components of the ORDENA system: (a) OD-Plot, (b) Trip Location, (c) Feature Importance, (d) Model Evaluation, and (e) Feature Detail views. The yellow-highlighted workflow illustrates four interaction levels, guiding the user through the exploration of OD flows.
• Figure 5: Case Study \ref{['subsec:caseA']} (A): OD-Plot visualization of car trips in Melbourne (Dataset A). Regular trips are represented by blue dots, while trips with fatalities are depicted in purple. Diagonal concentration indicates predominance of short-distance trips.