Introduction

The amount of data that needs to be analyzed is continually increasing. This presents constant challenges when it comes to selecting and running the most appropriate visualizations for each dataset, especially when working in contexts of large volumes of data, as Big Data is challenging analytical contexts, namely when aligning data and analytical requirements and using the most appropriate visualizations for supporting users to make more informed decisions.

In this context, the multiplicity of choices and the lack of clarity regarding analytical objectives make it difficult for users to establish effective connections between the two for data visualization [1]. Each dataset has unique characteristics and not all types of visualizations appropriately represent them [2]. Although some studies have proposed approaches to optimize data visualization ([3], [2], [1], [4], [5]), several challenges exist in aligning analytical requirements with the data, as well as translating domain questions, expressed in natural language, into analytical tasks. These tasks are then used to design analytical visualizations.

This paper proposes an assisted model-driven analytics approach to support analytical visualization, using domain knowledge and data as input. After mapping the data, the approach helps the user translate domain questions into analytical tasks that are supported by analytical visualizations. The proposed iterative process, from the identification of the most appropriate analytical tasks for each question to the analytical visualizations, is supported by the modeldriven analytics component of the approach. This component includes a Meta-model that contextualizes the data, the analytical tasks applied, and the analytical visualizations that can be used to analyze the results obtained from performing these tasks. This Meta-model formalizes the concepts needed to answer three fundamental questions: what, the type of data the user is dealing with; why, the reason why the user wants to analyze that data; and how, how the visualization is implemented in terms of design choice.

This paper is structured as follows. Section 2 presents related work. Section 3 presents the proposed model-driven assisted analysis approach. Section 4 presents the proposed Meta-model. Section 5 presents and discusses a demonstration case applied in the genomics domain. Finally, section 6 summarizes the conclusions and future work.

Related Work

In the context of research approaches to data analytics, model-driven approaches are presented through the concept of modeling real-world domains as a knowledge base to ease the analysis of the modeled domains. This type of approach generally focuses on facilitating visualization design choices but is not capable of bridging the mapping of domain data into visual channels [5]. We believe that within this space, it is possible to contribute towards the inclusion of conceptual models as domain knowledge that will be used to relate domain concepts with domain data and help translate user requirements. Some works of data analytic approaches, with a focus on modeling, such as [3], propose a model-driven architecture that allows automation for the creation of visualization through the translation of the user-specific objectives/goals. Other works, e.g. [2], intend to facilitate the design choices regarding visualizations to users who lack data analysis expertise through the use of a model-driven approach in which user requirements, data profiling, and visualization design are considered. In addition, other works use iterative goal-oriented models that specify visualizations to create dashboards [6] or propose visualization frameworks that map user requirements to data visualizations [1]. The work of [7] explores how joint interactive visualization can improve the communication of knowledge between different users, promoting mutual understanding through the visual representation of data.

These studies, although they share similarities with the work presented in this paper, do not provide sound mapping strategies between the required domain knowledge and data. There is a lack of an approach that guides the user in aligning analytical requirements with the data, as well as translating domain questions into analytical tasks to ensure that analytical visualizations adequately address the identified analytical tasks. This approach seeks to bridge this gap by aligning domain knowledge, domain data, and analytical requirements with suitable visualizations tailored to designed analytical tasks. Another defining mark of our approach is the support for identifying analytical tasks from analytical requirements using a taxonomy that maps user requirements into analytical tasks. For proposing the taxonomy, related works considered the works of [8,5], which describe visualization tasks at varying levels of abstraction and consider that analytical tasks are driven by the need to perform complex actions based on thorough data analysis ( [9,10]), and other low-level taxonomies ( [11,12]) that typically encompass simpler actions that do not require an in-depth analysis of the overall analytical context. This taxonomy is integrated into a Meta-model that contextualizes the domain questions, the analytical tasks and the analytical visualizations useful for decision-making.

Model-driven Analytics Approach

Data-oriented analytics guides the identification of valuable insights from vast amounts of data. This is particularly relevant in a context where Big Data imposes complex challenges in the alignment between analytical requirements and data. Usually, the application domain is described using a conceptual model, but the definition of the analytical requirements and the identification of the corresponding visualizations are often done by looking into the users' needs. This work aims to advance model-driven analytics with an approach that considers the domain knowledge and data as input and assists the user in translating domain questions expressed in natural language into specific analytical tasks that are supported by useful visualizations. The approach here proposed follows the human in the loop principle proposed by [5], being supported by an iterative analytical process to augment human capabilities (and not to replace them). As [5] highlights, this iterative process i) requires interactions between the user and the several analytical visualizations supporting many possible queries and, as such, handling complexity with data analysis at different levels of detail; ii) can be framed by three essential questions: what data the user is dealing with, why the user intends to use a visualization tool, and how the visual encoding and interaction are constructed in terms of design choices.

Despite the relevance of the proposal presented in [5], this does not describe the concepts of the domain or map the domain to data abstractions, focusing more directly on understanding the analytical tasks required to answer domain-specific questions and how visualizations can support these tasks. Our approach considers these three fundamental pillars presented by [5], what, why, and how. However, it also proposes an analytical approach that, besides dealing with the concepts of the domain and the data, guides the user in mapping the concepts of the domain with the data. This is essential to ensure alignment between the domain concepts and the available data, promoting a consistent and targeted analysis of the domain's analytical requirements. The approach here proposed (Figure 1) considers three main components:

• Domain Knowledge and Data, where a conceptual model of the domain is available describing the main concepts and relationships, as well as the available data and the domain questions for the data; • Model-driven Analytics, including the proposed Meta-model that contextualizes the data, the analytical tasks, and the analytical visualizations that can be used to analyze the results of the analytical tasks; • Assisted Model-driven Analytics, with four core steps guiding the proposed approach from the domain concepts, data, and questions to the visualizations. This encompasses the mapping of the domain concepts and data, the identification of the analytical tasks for the defined domain question(s), the processing of these tasks and, finally, the processing of visualizations that map the tasks' output into useful instruments for decision support.

Considering the Domain Concepts formalized in a data model (such as a Class Diagram or an Entity-Relationship Diagram) and a specific dataset for analysis (Domain Data), already with the prepared data, the first step of the assisted model-driven analytics components maps these two relevant pieces of information to check the alignment between them (Domain and Data Mapping). This step involves mapping the attributes defined in the classes of the data model with the attributes of the domain data available for analysis, ensuring a common understanding of the concepts and supporting data. A list, table, or another similar artefact must be made available as a result of this mapping step (Mapped Data). This information is useful for the identification of the analytical tasks (Analytical Tasks Identification), translating the Domain Questions (questions set by the domain user to be answered) into the Analytical Tasks that will be detailed with the help of the proposed Meta-model. This Meta-model includes a set of Analytical Tasks that define an iterative sequence of analysis processes. The Data Engineer plays a key role in supporting the identification of the analytical tasks needed to answer the domain's questions. These tasks are associated with output targets, which represent the analytical results (Analytical Outputs) obtained after the data analysis process (Analytical Tasks Processing). These are the inputs for the visualizations (Visualizations Processing). The Meta-model supports the identification of the appropriate visualizations according to the obtained results. This approach adopts a human in the loop philosophy, with the interaction of the Domain User and the Data Engineer and the processing components, and also interactions with or between components.

Model-driven Analytics Meta-model

The proposed approach is supported by a Meta-model that contextualizes the domain questions, the analytical tasks and the analytical visualizations useful for decision-making. This section first presents the proposed Meta-model and describes its main packages and concepts. The Unified Modeling Language (UML) Package Diagram presented in Figure 2 includes three main packages, What Dimension, Why Dimension and How Dimension, formalizing the concepts needed to answer three fundamental questions: what is the focus of the analysis?, why are we analysing these data?, and how can we analyse these data?. Each dimension includes its sub-packages and the dependencies associated with them. These dependencies can be classified into two types: import, where one package imports the functionality of another package, and access, where one package requires concepts or functionalities present in another package.

The What Dimension (Figure 3) package corresponds to the "what" component of the Metamodel and includes the Dataset sub-package with three detailed sub-packages: Items, Attributes and Data Types. Between these three sub-packages, there is an association between items and their respective attributes, and each attribute is associated with a specific data type.

The Why Dimension (Figure 4) includes three sub-packages: Domain Questions, Analytical Tasks and Targets. At the sub-package level, domain questions include the user's questions that are translated into analytical tasks and therefore require access to the functionalities present in that package. Furthermore, the analytical tasks sub-package requires access to the targets sub-package to filter, select or have as expected output one of the three possible targets available in the Meta-model, namely Attribute Target, Item Target and Dataset Target. The Analytical Tasks also have a connection with the Attributes sub-package as the Meta-model includes a relationship with a specific analytical task (Compute Attribute) which allows attributes to be derived. The other dependency occurs since certain analytical tasks allow for the creation of analytical visualizations that can be used to analyze the results of the analysis.

The How Dimension (Figure 5) package includes the Charts sub-package, with the set of and Spatial with the particular case of Geospatial data). The items included in these datasets aggregate different attributes that address simple data (such as a quantitative or qualitative value) or complex data (such as temporal or spatial data) and their corresponding values. Each attribute is associated with a specific data type. Datasets can include indexes for their items or attributes. Items can be classified and may establish relationships between them.

The "why" component, depicted in Figure 4, addresses the Domain Questions, which represent the questions the user wants answered and which can be translated into Analytical Tasks. The Analytical Tasks, which determine the actions that will be applied to the data and that can be formalized for addressing the analytical requirements of a domain, include tasks that express actions that can be used to find insights (tasks such as relationship, pattern, find extreme, find anomalies and find clusters), compare, determine distribution, organize, or to derive new data (that can be the expected output or be the input of another task). An analytical task usually selects data from a target (attribute, item or dataset), filters data from a target (attribute, item or dataset), and has as expected output a target (attribute, item or dataset). Depending on the The Meta-model establishes constraints on the types of charts that can be used for each analytical task since the decision on the chart will depend on the specific tasks that will be supported by analytical visualizations. The constraint next presented states that for the Relationship analytical task, whose objective is to identify and analyse the relationships and interactions between attributes, one of the possible visualization charts is a ScatterPlot (ChartType). The "how" component (Figure 5) addresses a set of analytical visualizations and their components, taking into account the analytical task(s) and the type of data used to meet users' analytical needs. Each visualization, represented by the Chart class, has derived attributes (nMarks, nAxis and nHeader) which are obtained from the number of associations between Chart and the corresponding components, ChartComponents. Each Chart can contain several chart components, depending on the type of chart (ChartType). In this way, the Chart class has different chart types (ChartType), characterized by the number of marks (nMarks), axes (nAxis) and headers (nHeader). The headers mentioned use the data from the corresponding attribute(s) to form a header with one or more entries and can be of type column or row; the axes use data that correlates with a range of values and can be of type x or y; and the marks control the type of MarkType, which can have different types of marks, such as color, size, text, This component presents constraints related to whether or not the ChartComponents can be included, as well as the number and use of each one, impacting the presentation of the final visualization. Each group of constraints has been formulated according to the analytical requirements needed to create each specific chart.

In terms of the relationships between the classes and components of the Meta-model, each target in the why component is linked to its respective class in the what component. In addition, each target can be associated with an attribute that contains a specific order in which it will be Due to the size of the images and space limitations in the paper, the global Meta-model can be found in1 , while the full list of constraints can be found in2 .

Demonstration Case: Genomics Domain

This section presents the application of the proposals to a demonstration case in the Genomics Domain. The domain concepts are formalized in the Conceptual Schema of Genome (CSG) [13], a data model expressed in a UML Class Diagram. Given the extension of this model, Figure 6 highlights the classes and relationships that are considered in this demonstration case. This model includes the Gene that is part of the ChromosomeElement, located in the Chromosome and can be transcribed as part of the TranscriptableElement. Additionally, a Chromosome can be located in several variations (Variation). This Variation class, Precise or Imprecise, can occur The domain data with a prepared dataset includes positions in the DNA (variants) where a variation may occur, the gene affected by each variant, and the genotype of the patient (one or two copies of the alternative allele). The dataset contains six columns representing this information: Chrom, POS, REF, ALT, Genotype, and Gene. Each variant is represented by its position in the DNA. This position is represented by the chromosome (Chrom), the sequence position where the variant occurs in the chromosome (POS), the reference allele (REF), and all possible alternative alleles that could be observed in that position (ALT). The Gene column defines the gene or genes affected by the studied variant. The genotype (Genotype) determines which alleles the patient has at the studied position. The reference allele (REF) is represented by a 0, and each of the alternative alleles is represented by 1, 2, 3, …. Since humans have two copies of each chromosome, the Genotype value also represents if the patient presents the variant in one copy (heterozygote) or both copies (homozygote).

The result of the first step of the proposed approach (Domain and Data Mapping) is shown in Table 1, mapping the domain concepts and the available data. In the example provided, there is a direct correspondence between the attributes defined in the classes of the conceptual data model and the attributes of the data available for analysis (Domain Data). For instance, the domain concept Chromosome.name, which is an attribute of type String in the conceptual data model, corresponds to the attribute Chrom in the dataset, which is also of type String and contains the actual values of the chromosome names. Similarly, the domain concept VariationPosition.start, which is an attribute of type Long Integer and indicates the starting position of a genetic variation, corresponds to the POS attribute in the available data, which is also a Long Integer and stores the positions of genetic variations. This mapping ensures that data types and attributes are compatible between the conceptual data model and the available data, and allows for data validation, checking that all defined concepts are represented in the available data.

Next, the object diagrams with the instantiation of the Meta-model are presented. Figure 7 presents the "Why Dimension", highlighting the domain question to be discussed, the analytical tasks needed to answer the question and the targets used by the analytical tasks.

For this dataset in VCF, Variant Call Format, the following domain question was formalized "What is the distribution of variants along the genome in a sample". In the second step of the For the "What Dimension", Figure 8, the first analytical task selects data from a set of data in a table. These table items integrate a set of attributes with a specific data type and the corresponding attribute values. The dataset includes indexes for items and attributes. The expected output target is a dataset with two new attributes (Left Allele and Right Allele) added to the initial dataset. These attributes have the condition that, depending on the position, whenever the allele is represented by 0, the derived attribute corresponds to the REF attribute, but if one of the alleles is represented by 1, 2, 3, among others, then the derived attribute corresponds to ALT, which represents all possible alternative alleles. In addition, the following analytical task involves selecting data from a table dataset, namely the dataset resulting from the previous task. The expected output target is a dataset with the six attributes belonging to the initial dataset (Chrom, REF, ALT, Genotype and Gene) and the two newly derived attributes (Left Allele and Right Allele). These analytical results derive from the processing of these analytical tasks belonging to the third stage of the approach, Analytical Tasks Processing.

After obtaining the analytical outputs, the Visualizations Processing step supports the development of a chart from the Determine Distribution task. Figure 9 shows the instantiation of the "How Dimension" component and the types of components required in terms of headers, axes and marks, as well as the type of ChartType used to form the analytical visualizations. The The visualization generated through the Tableau tool is of the GanttChart type (Figure 10), one of the possible charts according to the restrictions of the Meta-model and the type of data. This includes a header with the Chrom and Gene attributes, an axis corresponding to the POS attribute and three marks: the first text mark includes the Chrom, Gene, Left Allele and Right Allele attributes, the second shape mark includes to the POS attribute and, finally, the color mark highlights the Gene. The order of the attributes in the visual representation is determined by the hierarchy established in the data model (Domain Concepts). In this case, the Chrom belonging to the Chromosome class according to the data mapping is presented as the first attribute, followed by the Gene and then the numerical POS attribute. There is a hierarchy between the Chromosome class and the Gene class, and genes are constituent parts of these chromosomes. The POS attribute, being a numerical attribute associated with the horizontal axis of the chart, is found after the Gene attribute in the analytical visualization. Now, it is possible to analyze, for a given chromosome and gene, its position in the genome and the corresponding alleles. For example, the chr1 chromosome with the BRAC1 gene at a position between 20 and 30 million corresponds to the T/C alleles, with the Left Allele represented by T and the Right Allele represented by C.

The obtained visualization allows analytics for the defined domain question "What is the distribution of variants along the genome in a sample?". Although it is possible to create other types of charts, they would not be as effective in demonstrating what the Domain User requires. It is relevant to note that the visualization presented as a result of this demonstration case has been validated by a domain expert, thus ensuring that the selected visualization efficiently meets the analytical objectives. Furthermore, it is important to mention that it is up to the user to make the final decision regarding the choice of the most suitable visualization within the range of possible visualizations suggested in the Meta-model.

Based on this demonstration case, we found that the analytical approach proposed facilitates the development of visualizations that effectively address domain questions. By integrating domain knowledge and data as input, this approach aligns analytical requirements with data and assists users in translating domain questions into actionable analytical tasks supported by useful visualizations. The Meta-model plays a crucial role in this iterative process by contextualizing data, identifying applicable analytical tasks, and guiding the creation of visualizations. This approach is applicable across various domains and datasets.

Conclusions

In this paper, we have presented an approach to supporting analytical visualization that provides guidance, from mapping data and identifying analytical tasks to creating analytical visualizations capable of responding to users' analytical needs. This approach is supported by a Meta-model, which contextualizes the data, the analytical tasks and the analytical visualizations that make it possible to analyze the results of these tasks. To verify the validity of the approach, we applied it to a demonstration case in the genomics domain, presenting an example of a useful analytical visualization.

In future work, further evaluation and the extension of the Meta-model, to add interactivity between the user and the proposed visualizations, are considered.