Abstract
Standardizing data visualization in industrial settings is critical for the effective communication of machine learning insights. This paper introduces VisuOnto, a novel ontology designed to model the procedural steps of chart construction rather than just the final output. By formalizing the visualization pipeline—from canvas creation to data mapping and styling—VisuOnto enables the automatic generation of knowledge graphs that can be translated into executable code. This approach bridges the gap between high-level knowledge modeling and practical implementation, facilitating reproducibility and clarity in industrial data analysis.
Introduction
In industrial environments, machine learning (ML) is increasingly deployed to monitor complex processes such as robotic positioning and welding quality. However, the utility of these models depends on how clearly their results are communicated to stakeholders. Visualization is the primary tool for this communication, yet the industry lacks a standardized language to describe how these visualizations are built.
Currently, the logic behind a data scientist's choice of plot and the specific steps taken to create it are often lost. Existing ontologies are frequently too broad or focus solely on mathematical data properties, failing to capture the practical, procedural requirements of building industrial dashboards. This lack of standardization leads to misunderstandings and makes ML results difficult to replicate or discuss accurately.
Methodology
The researchers developed VisuOnto using the Human-Centered Collaborative Ontology Engineering Methodology (HCOME). This process was structured into four distinct stages to ensure both technical rigor and industrial relevance.
- Domain Analysis: The team studied existing visualization tasks at Bosch and analyzed popular libraries like matplotlib to identify the fundamental "grammar" of plotting.
- Concepts Formalization: Using Protégé and OWL 2, they defined classes and relationships, distinguishing between AtomicVisualisationTasks (basic components) and PracticalVisualisationTasks (complex procedures).
- Mechanism Investigation: The study explored how the ontology could serve as a template for Knowledge Graphs (KGs).
- System Deployment: The ontology was tested in real-world manufacturing environments to validate its utility.
Procedural Visualization Design
A core innovation of VisuOnto is treating visualization as a serialized process. Rather than defining a static object, the ontology models a "pipeline" of steps. This allows a complex industrial chart to be decomposed into a recipe of sequential actions: creating a canvas, defining subplots, mapping datasets to specific methods (e.g., line or scatter plots), and applying styles.
By aligning the ontology’s structure with common code libraries, the researchers created a system where high-level visualization requirements can be automatically encoded into a Knowledge Graph. This KG then serves as a blueprint that can be "decoded" back into procedural Python scripts.
Evaluation and Results
VisuOnto was evaluated using internal Bosch use cases from the automotive and IoT sectors, as well as practical requirements gathered from data scientists. Key results include:
- Automatic KG Generation: A developed GUI allows users to input requirements, which the system then encodes into a standardized Knowledge Graph.
- Code Decodability: In workshops, data scientists successfully translated these generated representations back into procedural Python scripts to complete analysis tasks.
- Requirement Fulfillment: The ontology met critical dimensions of Coverage (representing necessary charts), Procedure (capturing workflow), and Application (comprehensibility for industrial users).
Conclusion and Future Work
While VisuOnto successfully standardizes common industrial visualizations like scatter, line, and histograms, it remains an ongoing project. Future work will focus on expanding the variety of supported charts to include more niche industrial visualizations and increasing the level of automation. The ultimate goal is a seamless pipeline where the ontology can autonomously generate and execute visualization code with minimal human intervention, further streamlining the industrial data analysis workflow.
Summary updated