2.1. Application of Business Intelligence in Organizations

Organizations face a complex array of data and information, strong competition from direct and indirect competitors, and the need for rapid, correct decision-making. Data originates from various stakeholders, such as customers, suppliers, employees, information systems related to business transactions, facts, and external data sources. Organizations collect these data for intelligent decision-making. Hence, data and information become an organization’s strategic resources and the essential factor driving the company’s efficiency and effectiveness. Therefore, BI is an appropriate solution for organizations to analyze various data and as a decision-making tool (Olszak, 2020). The definition of BI is a system that combines architectures, databases, techniques, and analytics tools. The critical objectives of BI are to analyze historical and present data to create an interactive, dynamic view of an organization’s performance, data, and information for decision-makers (Sharda et al., 2018).

Klašnja-Milićević et al. (2017) presented a framework for BI processing using the Hadoop platform for data scientists. Similarly, Kumar and Belwal (2017) and Yulianto and Kasahara (2018) implemented BI in organizations using the Pentaho software, and Villegas-Ch et al. (2020) used traditional BI processes using the Weka tool in higher education institutions. Meanwhile, Fraihat et al. (2021) designed an architecture for and developed BI for the real-estate market using the Python platform, employing the extract, transform, load (ETL) process and Power BI software for data visualization. These frameworks are general BI tools and applications for organizations. They differ from the business indicator management system de Andrade and Sadaoui (2017) developed using the Joomla platform, that was integrated into the dashboard view of KPI monitoring. A little differently, Piri (2020) analyzed KPIs by interviewing the managers and adopted the KPIs dashboard design. As mentioned above, many tools such as Weka, Pentaho, Hadoop, and Power BI are the tools organizations use for BI applications.

2.2. Business Intelligence on Key Performance Indicators Context

Many researchers emphasize studying KPI in the BI process of strategic management. KPIs were adapted as part of business realization in BI. Data scientists have to understand indicator analysis before proceeding to data processing and data mining (Alaskar & Saba, 2020; Bréant et al., 2020; Khatibi et al., 2017; Ng et al., 2015; Pestana et al., 2020; Ren & Tao, 2012; Rocha et al., 2017; Visser, 2020). In the same way, Zoumpatianos et al. (2013) combined the KPI monitor processes: KPIs projection analysis, KPIs data calculation, expectation analysis, performance analysis, and impact analysis, with BI. Conversely, Maté et al. (2017) described a semantic approach for business analytics using KPIs. This approach begins with strategic modeling using a business intelligence model to manually build the KPIs relationships and relate them to the semantics of business vocabulary and business rules (SBVR) using the Java platform. The SBVR referred to as a business dictionary becomes the corpus for the KPI interpreter, which provides the meanings of KPIs to the online analytical processing interpreter. In another KPI study, Sultan et al. (2017) proposed a framework for detecting appropriate KPIs based on feature selection using Weka’s information gain algorithm and then determining the association rules adapted for the frequent pattern-growth algorithm and compact pattern tree (CP-tree).

Furthermore, Azzouz et al. (2020) designed four steps – namely, business context analysis, the definition of strategic objectives, multidimensional modeling, and system implementation – for strategic alignment of KPIs systems. Certainly, KPIs are the main focus when an organization needs to develop their BI for strategy monitoring and driving data performance. Therefore, KPIs are the key factors that must be calculated into the business insight process and integrated into all objective and goal projections.

2.3. Semantic Technology with Ontology for Key Performance Indicators

The definition of an ontology is a set of representational primitives used to model a domain of knowledge or discourse. The primitive representations for ontology are usually classes (or sets), attributes (or properties), and relationships (or relations between class members). Furthermore, these assets complicate the meanings and constraints on the rationally coherent application of information. In the context of database systems, ontologies can refer to the level of abstraction of data models. Moreover, an ontology is a hierarchical and relational model intended for knowledge modeling of individuals, including their attributes and relationships to other individuals. Typically, ontologies use language that abstracts data found in structures and schemes. More specifically, using an ontology provides a greater ability to express first-order logic than a database model allows. The ontology is the semantic level and the database schema is the physical level of the data model. There are several utilizations of ontologies to integrate heterogeneous databases, enabling interoperability among incongruent systems and services on a knowledge-based level, because ontologies operate independent of the lower level databases (Gruber, 2009).

There are some exciting applications of semantic technology related to KPIs. For example, Bai et al. (2014) applied ontology technology to KPIs in personnel performance evaluations. First, they created the evaluation ontology and instances. After that, they used semantic web rule language to determine the rules of the evaluation statement and then queried to retrieve the information using SPARQL Protocol and resource description framework (RDF) Query Language (SPARQL). Similarly, del Mar Roldán-García et al. (2021) proposed a semantic model related to the ontology concept which covered KPI relationships and adapted the ontology to information retrieval with SPARQL, based on domain expert validation. In contrast, Walzel et al. (2019) presented KPI ontologies as the knowledge base for systems engineering and machine setup in manufacturing systems. Özcan et al. (2021) introduced an ontology enrichment technique which matched domain-specific ontologies and cross-domain knowledge graphs for ontology creation in artificial intelligence applications. Furthermore, their technique supported the increased correct data entity identification. Consequently, the ontology concept adds to the value of KPI relationship studies, which can apply in various contexts.

2.4. Difference Between Ontology and Databases

Relational databases (RDB) rely on a schema for construction of a table, but ontologies depend on the knowledge base of various relationships between classes or data. RDB are limited to one type of relationship, for example, the foreign key. If the system wants to create a relationship, or join multiple tables in the database to search for something, it has only a foreign key. The semantic web can see multidimensional relationships, for example, inheritance, associated with, part of, and many other forms, containing logical relationships and restrictions. Systems can infer implicit information from ontologies, but databases cannot (Hebeler et al., 2009).

Databases mainly use close-world assumption; on the other hand, ontologies use open-world assumption (OWA) if the system includes incomplete information. This concept represents existing knowledge and suggests new information findings. For example, if there is no direct flight from one place to somewhere else, the result returned by the ontology will be “I do not know.” Furthermore, databases use unique name assumptions. The definition of this term declares that there is only one world available for one entity from the real world. Thus, in a database, numerous vital terms have to be defined. Furthermore, the difference between an ontology and a database is that a database does not use a taxonomy to represent an original aspect of an ontological reference, because databases are used to precisely and safely store large amounts of data. However, ontologies locate the application by combining semantic data, or communication between heterogeneous systems, and they share the knowledge and the structure of the information with humans or software (Sir et al., 2015).

Ontologies and databases can represent similar actualities, but the application of ontological technology can present advantages depending on the problem. Database technologies can efficiently manage massive amounts of data, whereas ontologies reasonably represent reality, but the technology is inefficient at managing instances when they are present in the format of an OWL or RDF file. When the number of instances increases, ontologies have to be stored in a database and provide an interface that supports access to the data. However, ontologies offer a restriction-free framework to denote that a machine can be a readable reality, even on the Web. This framework depends on an OWA, wherein information can be defined, shared, reused, or allocated. Furthermore, information can also be exchanged and used to make inferences or queries. Choosing one or the other technology depends on the end-user’s requirements. If the information needs to be shared on the Web, an ontology should provide a good solution. Nonetheless, the decision would probably include using both technologies when a large amount of data needs to be stored and properly managed. Ontologies provide an excellent way to characterize reality, but a database is certainly the better method for storing data when size is a consideration (Martinez-Cruz et al., 2012).

Some studies have explored database transformation from a RDB to an ontology file or a RDF. Zhao et al. (2019) created a schema-based graph mapping method to transform RDB schema and table relationships into a RDF data graph, and then used the SPARQL endpoint to query the RDB data in the RDF file. Additionally, Jun et al. (2020) improved the RDB to RDF method by solving two problems: duplicated data generation and semantic information loss. Finally, Devi et al. (2020) built an ontology system covering the information requirement and used RDB to RDF Mapping Language for RDB to RDF transformation.

2.5. String Similarity Search

String similarity mapping is an essential technique for data integration and searches. The problem with string searching is how to decrease the inconsistencies in origin data and search keywords and increase the tolerated accuracy rate. The string-based similarity method is divided into token-based similarity and character-based similarity methods. The token-based similarity method evaluates the similarity of two strings with the set form of the string being a set of tokens. This method compares the overlap of the members in the tokens sets. Examples of the token-based similarity method are Overlap, Jaccard, Cosine, Dice, and Fuzzy-Wuzzy token set ratios. The second method is the character-based similarity method, which compares the consistency of two strings by separating each string into a sequence of letters and measuring the edited distance. For two strings that are examined to be considered similar, the strings must have a similarity score that meets or exceeds a set threshold. Examples of this method are Hamming distance and Levenshtein distance (Angeles & Espino-Gamez, 2015; Appa Rao et al., 2018; Yu et al., 2016). Finally, there are methods that do not fall under string-based similarity methods, such as the Jaro-Winkler and Monge-Elkan methods, which employ different approaches for performing string similarity comparisons. Thus, those methods are referred to as hybrid similarity methods (Prasetya et al., 2018).

Angeles and Espino-Gamez (2015) reported that the accuracy rate of the Hamming distance approach was higher than that of the Jaro and Monge-Elkan approaches, but its overall performance was lower compared to the Jaro and Monge-Elkan approaches. Appa Rao et al. (2018) presented the Fuzzy-Wuzzy method as being more appropriate than Levenshtein distance and Sequence Matcher in regard to the string similarity matching of a mathematical formula. Using dynamic queries of the MySQL database, Rinartha et al. (2018) compared the efficiency of Jaccard similarity, MySQL pattern matching, Levenshtein distance, and MySQL Fulltext Index. They determined that the MySQL Fulltext Index was the most accurate method, while MySQL pattern matching was the fastest method for query suggestion.

3.1. Research Framework

The framework of mapping data schema with ontology (MapDS-Onto) is a part of our research about developing semantic BI. The MapDS-Onto framework, shown in Fig. 1, maps the factors of the KPIs formula with the associated column in a RDB using the relation of the data class and the related ontology. The MapDS-Onto framework will then create the factor dictionary used for BI generation.

Fig. 1

Mapping data schema with ontology framework. KPIs, key performance indicators.

The MapDS-Onto framework has four layers for objective processing. Layer 1 lists all of the databases in the organization’s information systems along with their different configurations and connections. Layer 2 is a database schema mapping (DaSMap) and begins with the experts (researcher and strategic administrator) helping to create the database structure model ontology (DSMO). The relationship structure of the DSMO is presented in Fig. 2. The DSMO is the model used for the DaSMap algorithm, and that applies the process from Zhao et al. (2019) for reading the database structures of all of the information systems and building the database structure ontology (DSO). Layer 3 is called the domain schema mapping (DoSMap). Here, experts analyze the organization’s KPIs attribute relationships and build the information forms to create the performance indicator ontology (PIO). This form receives the performance indicator property’s data from users. The PIO is built based on an adaptation of the process developed by del Mar Roldán-García et al. (2021). Then, the DoSMap algorithm inputs the DSO and PIO to begin its process. DoSMap is the mapping method connecting the DSO and PIO to allow searching the appropriate classes in the DSO with the formula’s factor of KPIs class in the PIO and building the factor dictionary file. The dictionary file serves as a compendium, implemented in Python, designated for storing data requisite for processing. Consequently, this dictionary file constitutes the relative column corpus of PIO, encompassing the column name, table name, and database name.

Fig. 2

Database structure model ontology.

3.2. Research Methodology

The research methodology of this study is divided into three phases as follows:

4.1. Case Study

PSU is recognized as the highest quality university in Thailand’s southern region by QS World University Rankings for 2021 (Quacquarelli Symonds, 2021). PSU has five campuses located in the provinces of Songkhla, Pattani, Phuket, Suratthani, and Trang. However, the PSU system does not have intelligent BI that supports decision-making for their executives. Therefore, using PSU as the case study, this research began by scoping two of their information systems: (1) PIS and (2) PRPM system. PIS presents the individual information of staff across the five campuses of PSU, encompassing a wide range of data such as educational records, employment history, job positions, salary and benefits, and leave information. For department managers, this system affords the capability to search and filter information pertaining to their subordinates. On the other hand, the PRPM system stands as the central repository for PSU’s research data, displaying individual research history, funding information, and other pertinent details. Furthermore, the system facilitates users in submitting research proposals to solicit funding from the university and record research data from the start to the end of the project. The MapDS-Onto framework was used to test the case study. After running the DaSMap process, DaSMap read the database structures of PIS and PRPM systems and generated the DSO that followed the pattern of the DSMO. The DSO graph in Fig. 7 shows the complex paths of the relationships between the classes and subclasses. For example, “total_budget_plan” is a subclass of column class and is a column in the project detail table. A related KPI for the two systems is the budget from external funding per person. The formula of this KPI is “sumif(budget, funding source)/countif(staff id, faculty name).”

Fig. 7

Database structure ontology graph.

From this KPI formula in the case study, the DoSMap algorithm separated the formula into four factors—total budget, funding source, staff ID, and faculty name—and matched their class string-similarities and created the factor dictionary. The factor dictionary from the DoSMap algorithm is written as follows:

According to the factor dictionary of a KPI, the budget from external funding per person has four factors in the KPI formula. An example of the implication of this dictionary is that the budget factor has to read the data from the “total_budget_plan” column in the “project_detail” table of the PRPM system. The semantic BI system will extract the related data following the path of the column name in the database to transform, load, calculate, and visualize data in the next action of our research.

Another example is a KPI about the number of foreign academic staff working full-time, which has the formula “sumif(academic staff, staff type, staff country).” The factor dictionary of this KPI is as follows:

From the factor dictionary of the above KPI, we point out the formula’s fit data path. The academic staff factor must read the “staff_id” column in the “personnel” table of the DSS system. The staff type factor must read the “staff_type_name” column in the “personnel” table of the DSS system. The staff country factor must read the “country_id” column in the “personnel” table of the DSS system.

In the “SelectedColumn” function and “SelectedClassDictionary” function, this research tested the reasonableness of the token-based and hybrid similarity methods and the hybrid similarity methods. The character-based similarity method was not applicable to this study, because that method calculates the differences in distance between each character in two words. However, in this case, the names of columns in the database were word groups. Hence, the other methods were better for string-similarity matching because the other methods are based on the word overlap technique.

Table 1 delineates the comparative results of matching accuracy and similarity scores between two categories of string similarity methods: (1) token-based similarity methods, including the Jaccard index, Sørensen-Dice, cosine similarity, overlap coefficient, and fuzz.token_set_ratio; and (2) hybrid similarity methods, namely Jaro-Winkler and Monge-Elkan. Upon considering the mean scores, it is discerned that token-based similarity methods yield a higher average score compared to hybrid similarity methods. Consequently, the token-based similarity method emerges as suitable for this case study. Looking more closely at the token-based methods, we see that the overlap coefficient and fuzz.token_set_ratio have similar scores.

Based on the exactitude comparison reported in Table 1, this research used the “fuzz.token_set_ratio” method for string-similarity matching. Although the “overlap coefficient” method produced the highest average score, the “overlap coefficient” method yielded the highest score values for more than one column. Therefore, it returned a false value. The “fuzz.token_set_ratio” method matched all the actual results and had an average score of about 0.76% lower than the “overlap coefficient” method. Based on evaluation of the practicability of application in this comparison, the “fuzz.token_set_ratio” method was more suitable for this research.

4.2. Discussion

When employing a specific BI software package, such as Weka, Pentaho, or Power BI, users need to understand how to select the required columns of data to run the BI process by themselves. This study eliminates this gap by developing a smart BI using semantic technology and providing user-friendly BI software for non-technical users. The previous works mainly applied ontology techniques to collecting and retrieving KPI values (Bai et al., 2014; del Mar Roldán-García et al., 2021; Maté et al., 2017). Some works used the ontology concept to create knowledge sharing between systems (Özcan et al., 2021; Walzel et al., 2019) by applying the ontology concept, data storage, and queries for information. However, this study proposes the ontology concept to store the database structures of different systems and share this information within the research framework, to promote suitable class matching and factor dictionary creation.

From the research question, “How can we decrease the complexity between non-technical user skills and database structure understanding using the ontology concept?”, when non-technical users use BI software for data visualization, they first must understand the data structure in the table form of a database. These users may be managers who need help understanding database concepts. In this situation, IT staff or data scientists become the responders. The decision-maker has no choice but to wait for the information from the responder. Therefore, the organization’s decision must take time (Alpar & Schulz, 2016; Lennerholt et al., 2018). The results of this study present the processes that go through database structure understanding before non-technical users work on the BI software. DSMO is the model used in the relationship learning process between the databases of various information systems. When the learning process is finished, we will save DSMO by renaming it as DSO. Hence, DSO is the instance of DSMO and uses it to share knowledge about the relationships of all databases in the information systems. DSMO and DSO will decrease the complexity between non-technical user skills and database structure understanding, because these users do not need to analyze the relationships of all databases. They can go through this complexity, and faster, to make decisions.

Many research studies have demonstrated the capacity for application systems to share information between systems through the concept of ontologies (Özcan et al., 2021; Walzel et al., 2019). Our research results, focusing on DSO, corroborate this assertion, as DSO serves as a center of knowledge regarding database relationships across multiple information systems. A significant finding is that the MapDS-Onto framework can be employed to match data across various information systems and provide appropriate data references to non-technical users. This framework utilizes DSO to familiarize itself with other systems in this environment and inform non-technical users about the most suitable data. As a result, these users do not need to understand how data is related within the database system.

Underpinning this framework are two pivotal algorithms, DaSMap and DoSMap. DaSMap algorithm employs DSMO and adapts the schema-based mapping to the RDF graph, as proposed by Zhao et al. (2019), thereby formulating a novel method for transforming database schema to the RDF. Central to the DoSMap algorithm is the method of string similarity comparison. The DoSMap algorithm reveals that token-based similarity methods, such as the “Fuzz.token_set_ratio” method, are more suitable than hybrid similarity methods. This finding aligns with the findings of Appa Rao et al. (2018), which posited that token-based similarity methods were more suitable than character-based methods for matching mathematical formulas. These results complement the research of Appa Rao et al. (2018), given that their study did not draw comparisons between the accuracy scores of token-based and hybrid similarity methods.

The three key findings of this study include: (1) DSMO, which serves as the model for creating DSO, a repository of knowledge about data relationships across multiple databases; (2) DaSMap and DoSMap, algorithms programmed within the MapDS-Onto framework; and (3) the “Fuzz.token_set_ratio” method, suitable for string similarity searching in this domain.

DSMO is a simple model for learning data relationships between the databases in the case study. An organization can reuse DSMO as a tool for creating data relationship knowledge and sharing this knowledge as a plugin on existing business software. The case study results guarantee the benefit of DSMO: creating automation through a neural framework. DSMO is converted to DSO in order to create a knowledge bank of data relationships. DSO is the knowledge of the data relationship that will be shared within the system. Many algorithms, as in Figs. 3-Fig. 4Fig. 56, present the programming process of data relationship learning and matching. The two main uses of this algorithm are DaSMap and DoSMap. The BI software developer can modify DaSMap to learn the data relationships in the different systems and adapt the DoSMap to match the suitable data in the search process. Furthermore, the “Fuzz.token_set_ratio” is the verified method that fits the string similarity comparison between the factor of the KPI formula and the data in database systems.

The benefits of this study from the point of view of non-technical users will happen after their existing BI system is improved. The many complexities of BI regarding database understanding will be replaced with the automatic matching process. The improved BI will be more user-friendly than in the past.

4.3. Limitations of This Study