The rapid growth in the volume, variety, and value of data has led to changing requirements for its storage, processing, and transmission. The problem of selecting data storage strategies that are optimally suited to the functional requirements and operational conditions is discussed in this paper. The article includes the comparison of 4 models of data representation: relational, non-relational, taxonomies, and knowledge graphs based on predefined criteria. To create a more holistic understanding of these models, an ecommerce example has been chosen. The mentioned models were examined using the Amazon Review dataset to demonstrate their advantages and disadvantages on real data.
Web applications are the backbone of the modern Internet, serving a wide range of purposes, including informational purposes, entertainment, education, e-commerce, etc. They accumulate a certain state, data of a different nature, and therefore, the task of storing such data is crucial. For the e-commerce industry, data storage is even more important, as this field produces a huge amount of information, which may contain the products sold on the marketplace, data on transactions performed by users, such as purchases, and much more. Many vendors provide the service in this field, and a much larger number of people use such services. Moreover, E-commerce hugely depends on information systems, particularly web applications that have databases as a foundation of data storage. Therefore, e-commerce is perfectly suited for comparing different data representation models used by such databases.
Considering a large set of different approaches to data storage and representation, the task of choosing a specific model is quite complex. This process includes understanding the key factors:
The nature of existing data.
The structure of the data, including the possibility of the data being semi-structured. The upstream and downstream of the data, i.e., the way the data enters a system and is further used.
The capabilities of the data storage, including validation, restrictions, or potential data size.
Thus, choosing the right model and data storage directly affects the work of the system, further challenges, and is ultimately important for business success. To understand the variety and the circumstances in which a specific approach arises, it is worth looking through the timeline of the models considered and their fundamental principles.
Relational model
One of the most common data representation models is relational. It is a first-order predicate
logic-based data management method introduced in 1969 by English computer scientist Edgar F.
Codd, that uses relational algebra as a foundation. All data are stored as tuples, grouped into
relations. Relational databases follow the relational model and organize data in tables with rows and
columns [
Non-relational model
Non-relational models have existed for a long time, since about the middle of the 20th century.
The concept was used as a separate term in the 90s of the 20th century to refer to databases that do
not use the relational model [
Taxonomy model
Taxonomies are a way to store data that is convenient for classification and categorization. The term was invented in 1813 by the Swiss botanist A. P. de Candolle and was used, among other things, for classifications in the natural sciences. This model organizes data hierarchically with parent-child relationships, so it is easy to create and understand hierarchy and convenient for classification tasks, yet it is not a good model to create relationships that do not fall into such connectivity.
Knowledge graph model
The knowledge graph (KG) is a flexible data model that uses a graph approach to describe
realworld concepts and their relationships in an organized structure. Nodes represent items like things,
people, or more general concepts, and edges show the connections between these entities,
representing the semantic (carrying meaning) relation between nodes, how different subjects relate
to each other, providing a technical basis and a convenient structure to work with knowledge [
Relational, non-relational, taxonomies, and knowledge graphs are the models that have existed for a long time, especially taking relational and non-relational approaches. As expected, there are many articles comparing these approaches based on various criteria.
Researchers often compare not exactly the models but rather specific implementations of
relational and non-relational approaches. The article [
Several articles [
Referring to the other two models: taxonomies and knowledge graphs, they are often used in pairs
for natural language processing and machine learning tasks [
E-commerce is also often used as an example to compare or demonstrate the application of these
models [
This article compares all 4 models on a selected example of an e-commerce system, identifying both the conceptual and the technical differences between these models, including the type and structure of data, purpose of the system, and organizational principles of each model, leading to a better understanding performance of such models.
Before moving on to an overview of the data models representing e-commerce data, it is useful to review these models from a more conceptual perspective, defining the technical and fundamental basis of the selected models and their features, without reference to specific data.The purpose of this comparison is to highlight the distinctive characteristics, advantages, and disadvantages of each model. The comparison considers 4 different data presentation models: relational, non-relational, taxonomy, and knowledge graph. The analysis is based on a set of predefined criteria depending on the common data management and knowledge representation requirements:
Structure – the fundamental organizational principle of data models Relations nature – how relationships between separate subjects are defined and managed Flexibility – the adaptability of the schema to evolving data models and handling different structures
Performance – efficiency in querying and data operations.
Comparison results according to these criteria were summarized in Table 1.
relationship-heavy and semantic queries
All the above characteristics define the foundations of these data models. As can be seen from the table, high performance and constraints of the relational model perfectly fit data warehouse tasks, including BI and data analytics. High-flexible non-relational models often form the basis of data lakes with large amounts of semi-structured data used for research or ML workloads. Hierarchical structure of taxonomy is widely used in medicine and biological sciences. In turn, knowledge graphs are well-suited for AI tasks, enabling semantic reasoning and the extraction of new knowledge from existing data. These characteristics lead to the specific advantages and disadvantages described in Table 2.
Easy to visualize and implement
Schema rigidity makes it less flexible for evolving data models Limited in representing semantic relationships beyond foreign keys
documents can be complex and inefficient Relations between subjects may be implicit or complex Often needs external tools for advanced analytics
or cross-category links Limited semantic depth, as it often lacks contextual information Inflexible when data does not fit neatly into a hierarchy
schema evolves dynamically Efficient traversal of complex relationships Excellent for reasoning, recommendations, and insights through inference
Having defined the fundamentals and key features of the selected data models, we can apply these models to represent e-commerce data. An important feature of data in this area is the presence of a large amount of heterogeneous data, which has a different structure and nature, sometimes semistructured and with complex relationships, thereby bringing functional requirements for the data models that represent such information.
To better understand the practical differences and strengths of the listed data models, we take a
real dataset containing e-commerce data and present its data using each model, which allows us to
understand the specific advantages and disadvantages of the models for this data. We use the
Amazon review data dataset (2018) [
The fields have the following description [
Using a relational model for such data, consider the following structure in Figure 2. It is worth noting that the metadata has nested data as values of certain fields (e.g., sales_rank, which is another JSON object), which is not a typical representation for the relational model, so this structure needs to be normalized and broken down into separate relations, as in the entity relationship diagram below, which form tables with foreign key relations according to Table 1.
Assessing the suitability of the relational model for such data, we can say that this approach ensures data consistency and allows us to detect the deviations of the data structure from the defined schema (schema anomalies) at the write step, but at the same time, this approach reduces the flexibility present in the metadata example. Possible changes in the data schema will be quite difficult to implement. implement. Such properties may be defined by the following advantages and disadvantages.
Advantages:
Well-normalized structure enforces integrity.
Optimized for structured queries (e.g., "Get all products under $5 by brand X").
Scalable for large data sets using joins and indexes.
Field asin title feature description
price image_url image_hd_url
also_buy also_viewed sales_rank
brand categories
Data type
Is always present array<string>
true (length varies) string string string float string string
Features must be stored as separate rows in a separate relation.
Querying relations (e.g., also_viewed, also_bought) requires JOINs.
Schema changes (adding new product fields) require migration. 4.2.
The JSON representation is natural for this model, so the structure in Figure 1 can be used without changes (Table 2).
The non-relational model is great for this kind of metadata structure, considering the possibility of nested fields and complex data structures, but this flexibility also has the downside represented by possible inconsistencies in the data schema. For example, incorrect or incomplete data can be stored, and thus, this approach requires more careful design of validation steps. Additionally, this model does not enforce clear and fixed relations (e.g., also_bought can have no direct relation to different products and can be just a list of strings), which negatively affects the performance of data operations. So, these features may be defined by the following advantages and disadvantages.
Advantages:
Schema flexibility (e.g., features, images, ranks, and lists can vary per product).
Easier to ingest and query with dynamic/optional fields.
Suited for full document retrieval (getProduct(“asin”)). true true true true true true array<string> array<string> false (length varies) false (length varies) object({string, int})
false (length varies) string array<array<string>>
true true (length and nesting vary)
Implicit relationships (e.g., also_bought is just a string list).
No referential integrity (risk of invalid asin references).
Complex aggregations or filtering documents are less efficient. 4.3.
Considering taxonomies and hierarchical approaches to data storage, this model is not suitable for storing the full metadata (most of the data fields have a non-hierarchical nature), but it is perfect for storing the product categories (Figure 3) and can be used to represent similar hierarchical data.
Taxonomy can be used as part of a search engine for filtering, or it can be used to store a category model to be referenced from other data models with the remaining metadata. For example, the product category taxonomy shown in Figure 3 can be stored as a separate model, providing quick and convenient access for other parts of the system that require this list of categories. Therefore, we can define the following advantages and disadvantages.
Advantages: Disadvantages:
Intuitive for classification-based navigation or filtering.
Easy to group/filter products for category pages.
Very lightweight structure.
This model is not suitable for representing features, price, or other metadata.
Not a complete data model – only works for hierarchical data. 4.4.
Knowledge graphs are not well-suited for storing a large amount of identical data, but they are great for displaying a domain of knowledge based on the data. The presented metadata suits as a good starting point for knowledge graph construction, as it contains a list of features of each product, which can be used to build a knowledge graph showing various features users can specify in a search query, and links to products that have such features. Figure 4 provides a simplified visualized subgraph of the knowledge graph. Particularly, this part describes the product “Girls Ballet Tutu”.
Considering the diagram, the product itself refers to additional descriptive metadata, the brand in this case, but there may be additional nodes. It's also important to note that, taking the “Coxlures” brand node as an example, it may have many other products with a similar subgraph, so these linked nodes do not exist just for the needs of only that product but are interlinked components of the full graph.
Within this subgraph, we can see another green subgraph, which represents the hierarchy of categories with which this product is associated. Nodes that represent similar products are marked in blue. And the red subgraph defines the features of this product that are directly important when using such a graph to make a recommendation based on a list of requirements.
Therefore, such a KG can be a good foundation for a recommender system to help a user choose a product based on a text description. This system can parse a search query into individual tokens containing a product description, features, and category, and then use this graph to identify products that have such characteristics. Examples of similar systems for different knowledge domains can be found across scientific publications [18-19].When modeling data in a knowledge graph, it may be necessary to define a structure in accordance with the knowledge domain – an ontology [20] that provides a shared vocabulary and logical structure for how entities relate to one another for a particular scenario.
The following list of advantages and disadvantages can be defined for such a case. Advantages:
The ability to obtain insights based on the data
Rich-typed semantic relations of the data subjects
Creating a knowledge graph based on the described data can be resource-consuming and require a complex procedure of acquisition
The article compares four ways to organize data: relational and non-relational models, taxonomies, and knowledge graphs. The focus lies on what structure and rules they use, how they are formalized and organized, and for which part of the data is best suited. To better demonstrate the practical application of these models, the article represents real e-commerce data with each model, defining possible scope for a specific model (e.g., representing complete data or only a subset), advantages, and disadvantages of this model.
The relational model is great for handling organized tabular data that has fixed relation constraints with good performance at scale. They work well for transaction systems and reporting. However, they may be inefficient with complicated relations and non-persistent data structures. Considering e-commerce data, this approach is well-suitable for representing large amounts of data of different nature and origin as long as the data have a consistent, permanent schema. Therefore, it naturally reduces flexibility and is not compatible with semi-structured data.
Non-relational models are flexible enough to keep semi-structured data in formats like JSON. This data model is quite good for storing data with a dynamic schema. However, they are less efficient and more implicit at handling complex relationships between records. This model can be a good choice for a variety of e-commerce data. As shown in the example, the data often has a different structure and contains nested fields, and a non-relational approach widely supports such data features. At the same time, this model requires more careful validation, and subsequent schema inconsistencies can be present and detected only during further data manipulation if validation was not implemented well. Another important point to consider is the performance of data operations, which may not be as high as for relational models.
Taxonomies are strict in terms of relation types and provide less flexibility for different tasks. They are perfect for classification tasks, but less suitable for everything else. This model can be used to sort or arrange material. However, their rigid hierarchical structure and limited ability to capture complex relationships (only parent-child relations) make them less suitable for dynamic or overlapping data. Speaking of the application in e-commerce, it is only suitable for hierarchical data (e.g., product categories), for which taxonomies are a good choice and can simplify the work with such hierarchies and improve storage efficiency.
Knowledge graphs naturally best represent knowledge areas and semantic relationships between subjects. Their graph-based structure provides the possibility to define dynamic schemas, enables semantic reasoning, and efficient traversal of complex relationships. Compared to other data models, an important advantage is the ability to define generic relationships between entities. The model is abstract enough to represent different areas of knowledge and the relation between related entities. It makes them uniquely suited for search engines with semantic search, machine learning, and decision systems. In e-commerce, this model is not used to store raw data, but rather to extract insights about the data based on their relationships, such as making recommendations with a list of requirements. Within these tasks, the model is essential, though more complex to implement and operate.
After all, modern integrated data systems use combinations of these models. Therefore, the best choice is a hybrid system architecture, where different data models are used at different stages of data processing and for different tasks within a single data processing system framework. Declaration on Generative AI
During the preparation of this work, the author(s) used ChatGPT in order to: Drafting content, Generate literature review and Citation management. After using these tool(s)/service(s), the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content. [18] Zhao, Xuejiao, et al. "Brain-inspired search engine assistant based on knowledge graph." IEEE
Transactions on Neural Networks and Learning Systems 34.8 (2021): 4386-4400. [19] Chicaiza, Janneth, and Priscila Valdiviezo-Diaz. "A comprehensive survey of knowledge graphbased recommender systems: Technologies, development, and contributions." Information 12.6 (2021): 232. [20] Lypak, O. H., Lytvyn, V., Lozynska, O., Vovnyanka, R., Bolyubash, Y., Rzheuskyi, A., & Dosyn, D. (2018, September). Formation of Efficient Pipeline Operation Procedures Based on Ontological Approach. In Conference on Computer Science and Information Technologies (pp. 571-581). Cham: Springer International Publishing.