Towards a new hybrid approach for building documentoriented data wareh

International Journal of Electrical and Computer Engineering (IJECE)
Vol. 12, No. 6, December 2022, pp. 6423~6431
ISSN: 2088-8708, DOI: 10.11591/ijece.v12i6.pp6423-6431  6423
Journal homepage: http://ijece.iaescore.com
Towards a new hybrid approach for building document-
oriented data warehouses
Nawfal El Moukhi, Ikram El Azami, Soufiane Hajbi
Department of Computer Science, Faculty of Sciences, Ibn Tofail University, Kenitra, Morocco
Article Info ABSTRACT
Article history:
Received Jul 18, 2021
Revised Aug 12, 2022
Accepted Aug 24, 2022
Schemaless databases offer a large storage capacity while guaranteeing high
performance in data processing. Unlike relational databases, which are rigid
and have shown their limitations in managing large amounts of data.
However, the absence of a well-defined schema and structure in not only
SQL (NoSQL) databases makes the use of data for decision analysis
purposes even more complex and difficult. In this paper, we propose an
original approach to build a document-oriented data warehouse from
unstructured data. The new approach follows a hybrid paradigm that
combines data analysis and user requirements analysis. The first data-driven
step exploits the fast and distributed processing of the spark engine to
generate a general schema for each collection in the database. The second
requirement-driven step consists of analyzing the semantics of the decisional
requirements expressed in natural language and mapping them to the
schemas of the collections. At the end of the process, a decisional schema is
generated in JavaScript object notation (JSON) format and the data loading
with the necessary transformations is performed.
Keywords:
Apache spark
Data warehouse
Document-oriented data base
hybrid approach
Natural language processing
Natural language toolkit
Not only SQL
This is an open access article under the CC BY-SA license.
Corresponding Author:
Nawfal El Moukhi
Department of Computer Science, Faculty of Sciences, Ibn Tofail University
University Campus, Kenitra, P.O. Box 133, Morocco
Email: elmoukhi.nawfal@gmail.com
1. INTRODUCTION
Since the 1980s, relational database management systems (DBMSs) have continued to grow in
importance compared to other data management systems. Today, still used by the majority of companies,
they are appreciated for their ability to ensure strong data consistency and guarantee reliability during
transactions through the atomic-consistent-isolated-durable (ACID) rules adopted by these systems [1].
However, the advent of the big data era and the explosion of data volumes to be processed have led
companies to rethink their data management models, especially as relational DBMSs have shown many
limitations in terms of distributed architecture. As a result, large companies such as Google, Amazon and
Facebook have embarked on several research projects that have given rise to several not only SQL (NoSQL)
database management systems, focusing on performance, reliability and consistency. These new systems are
characterized by the absence of a priori defined schema, which gives them great flexibility, and a distributed
architecture allowing horizontal scalability and a huge gain in processing time.
Today, NoSQL DBMSs are gaining in popularity [2], and there is no doubt that they will one day be
the most widely used systems in the world. Therefore, it has become necessary to rethink the design of data
warehouses to accommodate the new NoSQL model. The first research works [3]–[7] addressing this issue
were limited to developing methods to transform a relational data warehouse into a NoSQL data warehouse
by using model-driven architecture (MDA) rules and techniques. Although these methods are very important

 ISSN: 2088-8708
Int J Elec & Comp Eng, Vol. 12, No. 6, December 2022: 6423-6431
6424
for upgrading business intelligence systems to the new NoSQL logic, the issue of designing a data warehouse
from unstructured data sources still remains problematic. In this sense, researchers [8], [9] suggested to
transform the NoSQL database into a relational database. The latter can be used to generate a
multidimensional data warehouse using classical methods. Even if these methods have succeeded in
integrating unstructured data sources, they do not allow to exploit the performance of NoSQL systems and
may have many limitations when dealing with large amounts of data. Other research works [10]–[13] have
used the MapReduce framework proposed by Google in 2004 [14] to generate a NoSQL data warehouse.
These methods represent a great progress in the field, since they allow to solve the problem while remaining
in the NoSQL logic throughout the design process. However, the MapReduce framework has several
limitations [15] and the paradigm (data-driven) adopted by these methods neglects the users’ requirements,
hence the interest of the present work.
In this paper, we introduce a new hybrid method that allows to generate a NoSQL data warehouse
by considering both user requirements and data sources thus adopting the best paradigm to design a data
warehouse [16]. The new approach uses a document-oriented database as a data source and allows first to
analyze the different JSON files in order to detect the structure of each collection and to extract its schema.
The second phase consists in mapping decisional needs expressed in natural language to the structures of the
collections using natural language processing techniques. At the end of the process, a multidimensional
model is generated and integrated into a NoSQL data warehouse for each expressed need. The paper is
organized: in section 2, we present the main steps of the new method. In section 3, we use a real database to
test the new method and discuss the results. Section 4 is devoted to draw conclusions and present research
perspectives.
2. RESEARCH METHOD
In this paper, we propose a new hybrid method for generating a data warehouse that can meet the
decision-making needs of decision-makers. In the first phase, the new method exploits the speed of the spark
framework-which is much faster than MapReduce according to several research works [17]–[26] in analyzing
large amounts of unstructured and distributed data, to generate a general schema for each collection. This
allows to extract the structure of a large amount of data in a reasonable time, thus revealing the richness of
the data stored in a document-oriented database. In the second phase, decisional needs expressed in natural
language by the decision-makers are semantically analyzed and compared with the different general schemas
extracted by the first data-driven phase. Such a mapping allows to identify the collection that best meets a
decisional need, and to select only those fields that are useful for the user. Figure 1 shows the main steps of
the new hybrid approach.
As shown in the figure, the new hybrid approach consists of 4 main steps:
a. Schemas extraction: a document-oriented database is composed of several collections, and each collection
contains JavaScript object notation (JSON) (or XML) documents distributed across multiple partitions.
This first step consists in exploring and reading the JSON files of a collection, contained in each partition.
At the end of the step, the schema (DataFrame) of each partition of the collection is extracted, thus
obtaining a multitude of schemas for the same collection.
b. Merging schemas: since we are working on a NoSQL database and the source data is unstructured, the
schemas of the various partitions are different and have common fields and new fields. Therefore, it is
essential to merge the schemas to get a first version of the general schema of the collection. Thus, the
common fields must be retained once and the new fields added. A simple merging of the different
extracted schemas will generate errors for the following reasons: i) some fields do not have the same data
types, ii) there is a significant change in the structure of some fields, and iii) the union() method can only
merge schemas with the same structure. To successfully merge the schemas, we first need to convert the
columns to strings to make sure the data types are compatible. Since JSON RDD allows the union of
schemas even when the structures are different, we need to convert the DataFrame to JSON RDD using
the toJSON() method, before uniting the partition schemas. In this way, we avoid the second and third
errors.
c. Merging fields: The first version of the general schema is the result of merging the schemas of the
partitions. As a result, identical fields with different names are also retained. This step remedies this
problem by analyzing the semantics of the names, using the NLTK library of python, and comparing
them for similarities. The process of merging fields is done in stages and by tree level of the general
schema. Thus, the fields at the highest level of the tree structure are analyzed and compared first. Fields
with a score of 50% or more are proposed to the designer, and only after having made the necessary
mergers that the program move on to the second level and so on, until the last level of the tree. Figure 2
shows the process of merging fields.

Int J Elec & Comp Eng ISSN: 2088-8708 
Towards a new hybrid approach for building document-oriented data warehouses (Nawfal El Moukhi)
6425
Figure 1. The new hybrid approach process
Figure 2. The new hybrid approach process
d. Semantic analysis and mappings: In this step, a decisional need expressed in natural language is
compared semantically to the general schemas of the collections. In order to perform this comparison, a
collection sentence is generated from each general schema by concatenating the collection name and the
field names. Thus, the decisional need and the collections sentences are semantically analyzed and
compared and each comparison results in a similarity score. Only collections sentences with a score
higher than or equal to 50% are retained and the corresponding general schemas are proposed to the user.
The general schema proposed by the system and selected by the user will enable him, on the one hand, to
check the availability of the data he needs for decision making, and on the other hand, to discover other
fields likely to interest him and to meet a latent and unexpressed decision-making need. After selecting
the relevant fields, a decisional schema that meets the user’s need is generated in JSON format and this
schema is used to load the data. When loading the data, the necessary operations induced by the step of
merging fields, must be applied. Thus, for the merged fields, the data of the non-retained fields must be
loaded into the retained ones. At the end of the process, we obtain a version of the collection intended
specifically for decision-making use and meeting a specific decisional need. All the collections that meet
the decisional needs expressed by the users will form our NoSQL data warehouse. Figure 3(a) shows the
algorithm of the first data-driven phase and Figure 3(b) the algorithm of the second requirements-driven
phase.

 ISSN: 2088-8708
6426
(a)
(b)
Figure 3. Algorithm of the new hybrid approach (a) algorithm 1: the first data-driven phase, and
(b) algorithm 2: the second requirements-driven phase
3. RESULTS AND DISCUSSION
In order to illustrate each step of the new approach, we used a real-world example of a sales
collection containing various data and distributed across three partitions. Obviously, the document is the
central concept of a document-oriented database, which encapsulates and encodes its data in some standard
format. The most widely adopted format is JSON, which we use as a reference in this work. Figure 4 shows a
document from the “Sales” collection contained in partition 1. The document is organized according to two
nesting levels: i) the first level contains information about a sale made at a given time. In addition to the
“date” and time, it contains data about the “amount” paid by the customer and the “quantity” purchased and
ii) the second level contains two main fields: "store" containing information about the store that made the
sale, and “customer” containing information about the customer who bought from the store.
The extraction of the collection schema first involves the extraction of the schema of each partition
of the collection. Figure 5 shows the extracted schemas of the “Sales” collection with the schema of partition
1 in Figure 5(a), the schema of partition 2 in Figure 5(b) and the schema of partition 3 in Figure 5(c). After
merging the schemas, we will get the schema in Figure 6 containing all the fields with the dataframe format
in Figure 6(a), and the JSON format in Figure 6(b). Table 1 presents the similar fields proposed by the
system for a possible merger.

6427
Figure 4. Example of a document from the “Sales” collection
(a) (b) (c)
Figure 5. The schemas of partitions of the “Sales” collection (a) schema of partition 1, (b) schema of partition
2, and (c) schema of partition 3
In total, the system proposes to the user to merge 26 fields, and 16 fields are to be merged (61.5% of
the total proposed). The fields “employee number”/”employee phone”, “country”/”country code”, “country
code”/”Postal code”, “country code”/”store country”, “store city”/”store country” refer to different
information despite the similarity calculated. For the 2 fields “first name” and “last name”, they must be
merged into a single “full name” field. When loading the data, a concatenation of the data of the fields “first
name” and “last name” should be considered. Figure 7 shows the second general schema of the collection in
JSON format, after performing the necessary mergers.

 ISSN: 2088-8708
6428
(a) (b)
Figure 6. The 1st
general schema in (a) dataframe format and (b) JSON format
Figure 7. 2nd
general schema in JSON format

6429
Table 1. The fields proposed for merging after the semantic comparison operation
Field 1 Field 2
purchaser customer
country nation
employee number employee phone
first name full name
first name last name
full name last name
address address details
city store city
country country code
country store country
country code postal code
country code store country
store city store country
We assume that the user has expressed the following decisional need: “We want to analyze the
quantity sold by date, by store, by city and by country”. The combination of our “sales” collection gives the
following result: “sales, amount, date, quantity, customer, address, city, country, street, type, age, name,
employee, employee number, employee phone, full name, store, address, country code, postal code, store city,
store country, number, phone”. We also assume that we have two other collections with the following
combinations: i) collection “suppliers”: suppliers, number, name, phone, address, city, country, cost, quality,
date, delivery number, mode, total demand; ii) collection “Products”: products, number, name, Reference,
Description, color, type, width, height, weight, production material, popularity, quality, price. Table 2
summarizes the results of the semantic comparisons between the expressed need and the 3 collections. At the
end of the process, a decision schema that meets the user’s need is generated. Figure 8 shows this decision
schema in JSON format.
Table 2. scores of the semantic comparison between the need expressed by the user and the collections
Sales Suppliers Products
"We want to analyze the quantity sold by date, by store, by city and by country" 0.733 0.527 0.147
Figure 8. Decisional schema in JSON format
4. CONCLUSION
In this paper, we presented a new hybrid approach to generate a document-oriented data warehouse
from unstructured data sources. The new approach is not limited to the analysis of data sources but also
includes the analysis of needs, hence its hybrid nature. The semantic analysis of the decisional need and its
comparison with the general schemas allows the system to assist the user in the choice of the collection and
the fields to be retained in the final decisional schema. In this way, only data useful for decision-making will
be loaded into the data warehouse and useless data will be eliminated. The prior knowledge of the structure

 ISSN: 2088-8708
6430
of the decision schema and the tree structure of the fields greatly facilitates the formulation of decision
queries in the document-oriented data warehouse. Our research perspectives consist in integrating other types
of NoSQL data sources such as column-oriented databases or key/value-oriented databases.
REFERENCES
[1] S. Shetty, B. D. Rao, and S. Prabhu, “Growth of relational model: interdependence and complementary to big data,” International
Journal of Electrical and Computer Engineering (IJECE), vol. 11, no. 2, pp. 1780–1795, Apr. 2021, doi:
10.11591/ijece.v11i2.pp1780-1795.
[2] “DBMS popularity broken down by database model Number of systems per category, December 2020.” 2021.
[3] F. Abdelhedi, A. Ait Brahim, F. Atigui, and G. Zurfluh, “MDA-based approach for NoSQL databases modelling,” in Big Data
Analytics and Knowledge Discovery, Springer International Publishing, 2017, pp. 88–102., doi: 10.1007/978-3-319-64283-3_7.
[4] M. Chevalier, M. El Malki, A. Kopliku, O. Teste, and R. Tournier, “Implementation of multidimensional databases with
document-oriented NoSQL,” in Big Data Analytics and Knowledge Discovery, Springer International Publishing, 2015, pp. 379–
390., doi: 10.1007/978-3-319-22729-0_29.
[5] M. Y. Santos, B. Martinho, and C. Costa, “Modelling and implementing big data warehouses for decision support,” Journal of
Management Analytics, vol. 4, no. 2, pp. 111–129, Apr. 2017, doi: 10.1080/23270012.2017.1304292.
[6] F. Abdelhedi, A. A. Brahim, and G. Zurfluh, “Formalizing the mapping of UML conceptual schemas to column-oriented
databases,” International Journal of Data Warehousing and Mining, vol. 14, no. 3, pp. 44–68, Jul. 2018, doi:
10.4018/IJDWM.2018070103.
[7] K. Dehdouh, F. Bentayeb, O. Boussaid, and N. Kabachi, “Using the column oriented NoSQL model for implementing big data
warehouses,” 2015.
[8] Z. Aftab, W. Iqbal, K. M. Almustafa, F. Bukhari, and M. Abdullah, “Automatic NoSQL to relational database transformation with
dynamic schema mapping,” Scientific Programming, vol. 2020, pp. 1–13, Jul. 2020, doi: 10.1155/2020/8813350.
[9] B. Maity, A. Acharya, T. Goto, and S. Sen, “A framework to convert NoSQL to relational model,” in Proceedings of the 6th
ACM/ACIS International Conference on Applied Computing and Information Technology-ACIT 2018, 2018, pp. 1–6., doi:
10.1145/3265007.3265011.
[10] S. Bouaziz, A. Nabli, and F. Gargouri, “Design a data warehouse schema from document-oriented database,” Procedia Computer
Science, vol. 159, pp. 221–230, 2019, doi: 10.1016/j.procs.2019.09.177.
[11] K. Ma and B. Yang, “Introducing extreme data storage middleware of schema-free document stores using MapReduce,”
International Journal of Ad Hoc and Ubiquitous Computing, vol. 20, no. 4, 2015, doi: 10.1504/IJAHUC.2015.073439.
[12] H. Chai, G. Wu, and Y. Zhao, “A document-based data warehousing approach for large scale data mining,” in Pervasive
Computing and the Networked World, Springer Berlin Heidelberg, 2013, pp. 69–81., doi: 10.1007/978-3-642-37015-1_7.
[13] K. Dehdouh, O. Boussaid, and F. Bentayeb, “Big data warehouse,” International Journal of Decision Support System Technology,
vol. 12, no. 1, pp. 1–24, Jan. 2020, doi: 10.4018/IJDSST.2020010101.
[14] M. Dayalan, “MapReduce: simplified data processing on large cluster,” International Journal of Research and Engineering, vol.
5, no. 5, pp. 399–403, Apr. 2018, doi: 10.21276/ijre.2018.5.5.4.
[15] S. A. Thanekar, K. Subrahmanyam, and A. B. Bagwan, “Big data and MapReduce challenges, opportunities and trends,”
International Journal of Electrical and Computer Engineering (IJECE), vol. 6, no. 6, pp. 2911–2919, Dec. 2016, doi:
10.11591/ijece.v6i6.10555.
[16] N. El Moukhi, I. El, A. Mouloudi, and A. Elmounadi, “Merge of X-ETL and XCube towards a standard hybrid method for
designing data warehouses,” International Journal of Advanced Computer Science and Applications, vol. 10, no. 10, 2019, doi:
10.14569/IJACSA.2019.0101019.
[17] Z. Ruihong and H. Zhihua, “WITHDRAWN: comparative research on active learning of big aata based on mapreduce and spark,”
Microprocessors and Microsystems, Nov. 2020, doi: 10.1016/j.micpro.2020.103425.
[18] X. Wu and S. Ji, “Comparative study on MapReduce and spark for big data analytics,” Journal of Software, vol. 29, no. 6, pp.
1770–1791, 2018
[19] E. P. S. Castro, T. D. Maia, M. R. Pereira, A. A. A. Esmin, and D. A. Pereira, “Review and comparison of Apriori algorithm
implementations on Hadoop-MapReduce and Spark,” The Knowledge Engineering Review, vol. 33, Jul. 2018, doi:
10.1017/S0269888918000127.
[20] J. Shi et al., “Clash of the titans,” Proceedings of the VLDB Endowment, vol. 8, no. 13, pp. 2110–2121, Sep. 2015, doi:
10.14778/2831360.2831365.
[21] N. Ahmed, A. L. C. Barczak, T. Susnjak, and M. A. Rashid, “A comprehensive performance analysis of apache hadoop and
apache spark for large scale data sets using HiBench,” Journal of Big Data, vol. 7, no. 1, p. 110, Dec. 2020, doi: 10.1186/s40537-
020-00388-5.
[22] I. Mavridis and H. Karatza, “Performance evaluation of cloud-based log file analysis with apache hadoop and apache spark,”
Journal of Systems and Software, vol. 125, pp. 133–151, Mar. 2017, doi: 10.1016/j.jss.2016.11.037.
[23] A. Mostafaeipour, A. Jahangard Rafsanjani, M. Ahmadi, and J. Arockia Dhanraj, “Investigating the performance of Hadoop and
Spark platforms on machine learning algorithms,” The Journal of Supercomputing, vol. 77, no. 2, pp. 1273–1300, Feb. 2021, doi:
10.1007/s11227-020-03328-5.
[24] J. Veiga, R. R. Exposito, X. C. Pardo, G. L. Taboada, and J. Tourifio, “Performance evaluation of big data frameworks for large-
scale data analytics,” in 2016 IEEE International Conference on Big Data (Big Data), Dec. 2016, pp. 424–431., doi:
10.1109/BigData.2016.7840633.
[25] Y. Benlachimi, A. El, and M. Lahcen, “A comparative analysis of hadoop and spark frameworks using word count algorithm,”
International Journal of Advanced Computer Science and Applications, vol. 12, no. 4, 2021, doi:
10.14569/IJACSA.2021.0120495.
[26] E. L. Lydia, R. M. Vidhyavathi, I. Pustokhina, and D. A. Pustokhin, “Comparative performance analysis of apache spark and map
reduce using k-means,” International Journal on Emerging Technologies, vol. 11, no. 2, pp. 198–204, 2020,

6431
BIOGRAPHIES OF AUTHORS
Nawfal El Moukhi was born in Salé in 1987. He graduated from the School of
Information Science of Rabat and he got a Master’s degree in applied computer science from
Mohammed V University in 2013. He holds a PhD in computer science from Ibn Tofail
University in Kenitra, Morocco. His research interests are data warehousing, data mining and
big data. He can be contacted at email: elmoukhi.nawfal@mail.com.
Ikram El Azami received his PhD degree from the University of Valenciennes
and Hainaux Combrésis, France, and The University of Sidi Mohammed Ben Abdellah, Fez,
Morocco. He is currently full professor in the Department of Computer Science at the Faculty
of Sciences, Ibn Tofail University of Kenitra. He’s a member of the MIS (Multimedia and
Information Systems) research Team of MISC Laboratory (Communication Systems and
Information Modeling), Faculty of Sciences of Ibn Tofail University. His research interests are
information systems engineering, business intelligence, big data and cloud computing. He can
be contacted at email: akram_elazami@yahoo.fr.
Soufiane Hajbi is a computer engineer. Currently, he is a PhD student at Ibn
Tofail University in Kenitra, Morocco. His research interests are natural language
processing, data warehousing, data mining and big data. He can be contacted at email:
soufiane.hajbi@uit.ac.ma.

Towards a new hybrid approach for building documentoriented data wareh

More Related Content

Similar to Towards a new hybrid approach for building documentoriented data wareh (20)

More from IJECEIAES (20)

Recently uploaded (20)

Towards a new hybrid approach for building documentoriented data wareh