Introduction

Visual analytics has been essential for data analysis in a wide range of applications in modern industries, such as in exploratory data analysis for gaining understanding and first insights for the data [1,2,3], presenting to the operators for identification of interesting data snippets [4], and representing the results of machine learning analytics [5,6]. However, there has been limited research of semantic technology on visual analytics, especially in industry [7]. This hamstrings the transparency and reusability, especially for non-data-scientists users, such as engineers, managers, who are important stakeholders in multi-disciplinary industrial projects [8].

Moreover, existing ontologies [9] and tools discussed the general purpose data visualisation, but limitedly exploited domain knowledge (such as feature type information of the data) or the essential procedures for visual analytical scripts [10]. To this end, we propose a semantic system that relies on semantic technologies for user-friendly visual analytics, offering domain knowledge supported visual analytics with no-coding experience. Our approach encodes visual analytical solutions in reusable knowledge graphs (referred to as visual KG) via graphic user interface (GUI) and reasoning, which can be translated to executable scripts [11]. Further more, our approach incorporates domain knowledge in a type of automatic manufacturing, automated welding by linking domain ontologies and visual ontology (a set of axioms that encode domain knowledge of visual analytics, such as visualisation methods, procedure, and constraints) [12,13]. In automated welding, robots press metal car body pieces together and a current flows through the robot electrode and the car bodies to melt the metal materials, which generates a welding nugget that connect the metal pieces [14]. This poster paper presents preliminary evaluation of our approach with a Bosch use case which shows promising results. This paper extends the visual KG part of our ISWC full paper about executable KGs [15] by giving more examples and technical details of KG construction and evaluation.

Our Approach

System Overview. We now walk through our system (Figure 1a). From the bottom left, the Non-KG Data can be e.g., relational tables (csv), json files. For simplicity, we focus on relational tables. These data are mapped to the domain ontology 𝑂 𝑑𝑜 via Data-to-Domain (Data2Domain) Mapping. The Data2Domain Mapping maps tables and table columns to classes in the 𝑂 𝑑𝑜 . Through Data Integration, domain KGs are generated. The Visual Ontology 𝑂 𝑣𝑖𝑠𝑢 contains axioms that describe the knowledge for visual analytics. The Visual KG Construction follows the schema defined by 𝑂 𝑣𝑖𝑠𝑢 and generates Visual KG [16,17], either manually [18] by a GUI or automatically following several Visual KG templates (also stored in 𝑂 𝑣𝑖𝑠𝑢 ). The generated Visual KG can be stored in Visual KG Catalogue for later reuse. The Visual KG Execution translates Visual KG to executable scripts and run the scripts, resulting the Visualisation Plots [19,20].

The Data2Domain Mapping is provided by domain experts in the form of rules as 𝑞 sql (𝑅) → 𝐴, where 𝑞 sql (𝑅) is an SQL query on table R that returns table columns and A is a class in the ontologies. We choose OWL 2 QL for 𝑂 𝑑𝑜 becasue it is optimised for efficient query answering over relational databases. We choose OWL 2 EL for 𝑂 𝑣𝑖𝑠𝑢 because of its expressivity and it is still polynomial for query answering. Between Data2Domain Mapping and 𝑂 𝑑𝑜 there exist class links, since Data2Domain Mapping maps the Non-KG Data to classes in 𝑂 𝑑𝑜 . All classes in 𝑂 𝑑𝑜 that represent features in data are linked to classes in 𝑂 𝑣𝑖𝑠𝑢 via the subclass axioms. Example 3. Take Fig. 2 as an example. Once the users choose to create a visual KG, the GUI will use the template in Figure 1b, which contains several owl:NamedIndividuals of the types Visu-alPipeline, CanvasTask, PlotTask, and DescriptionTask. Next, the users will need to select the input data, and add several entities of PlotTasks from available tasks based on 𝑂 𝑣𝑖𝑠𝑢 . For each PlotTask, the input data, the method and some parameters are mandatory to be given. The users need to configure the PlotTasks by specifying e.g., the inputs, line colour, line width. After that, they can also configure the CanvasTask and the DescriptionTask by giving the x label, y label, legend, etc.

Example 4. Our approach first identifies there exists three input features with the labels: target, estimated training and estimated test. Then it generates a KG from the template in Figure 1b with three entities of PlotTasks between the CanvasTask and DescriptionTask.

All of them are subclasses of TimeSeries. For TimeSeries, our approach uses the rule that adopts LinePlotMethod as the recommended plot method and adds its typing information: ∃hasMethod − .(Task ⊓ ∃hasInput.TimeSeries) ⊑ LinePlotMethod Visual KG Execution. The execution of visual KGs is language-dependent. Here we use Python for discussion: each individual of PlotMethod is a Python script for plotting, whose mandatory inputs/outputs and parameters are clearly defined. Each Visual KG contains an entity of VisualPipeline and a series of Tasks connected with hasNextTask. Thus, the execution of a Visual KG invokes the Python function scripts with the inputs/outputs and parameters given by Data and datatype properties of KGs, according to the order defined by hasNextTask.

Evaluation and Conclusion

Industrial Use Case. We tested our approach in an industrial scenario of machine learning based quality monitoring for automated resistance spot welding at Bosch, which is an impactful welding process accounting for the production of over 50 million cars globally every year. We invited industrial users from Bosch, including non-data-scientists to workshops for evaluation.

The visual KGs ran on a sample welding production dataset collected from a factory in Germany, contains 4585 welding operation records.

Visual Analytical Tasks and Reusability. The tasks in the use case can be categorised into two types: I. Data inspection (Figure 2a and Example 3), and II. Results visualisation. Type I tasks are to inspect various data in exploratory data analysis for understanding the data, and Type II tasks are to visualise results of statistical/ML analytics for discussing and interpreting the statistical/ML models for decision-making.

Example 5. For a new task of Type II and a new dataset in Figure 3a, the users can reuse the VisualPipeline in Figure 2a by simply modifying the input data entities (1), the plot methods (2), and properties of the plotting tasks (3, e.g., colour, marker size, labels, legend). This demonstrates the good reusability of our approach.

Transparency and Coverage.

We programmatically generated 242 KGs and use this as the Visual KG Catalogue. We organised extensive workshops and collected 24 reports from Bosch welding experts, engineers, semantic experts, data scientists. They answered questions such as "I found the semantic system helps to improve the transparency of visual analytics compared to the case without the approach", and gave scores ranging 1-5 (from disagree, fairly agree, neural, fairly agree, to agree) which aggregated to 4.28 ± 0.47 (mean ± standard deviation) for the transparency. After discussion, we categorised most tasks of visual encountered in our project in groups (Table 1), and give the coverage percentage according to our empirical cases. We can see that most of categories can be covered (above 80%) Conclusion and Outlook. This poster presents our system of domain knowledge supported KG construction for visual analytics, which offers good reusablity, transparency and coverage for the visual analytic tasks. We evaluated the system on a Bosch welding use case with promising results. As future work we plan to study hierarchical topic modelling to better organise our visual KG catalogue and push the deployment further. We also plan to further improve the system and organise demonstrations to a broader audience [21].