2021 IEEE 25th International Enterprise Distributed Object Computing Workshop (EDOCW)

Enterprise Architecture based Representation of

2021 IEEE 25th International Enterprise Distributed Object Computing Workshop (EDOCW) | 978-1-6654-4488-0/21/$31.00 ©2021 IEEE | DOI: 10.1109/EDOCW52865.2021.00055

Architecture and Design Patterns for Machine

Learning Systems
Hironori Takeuchi Takuo Doi Hironori Washizaki Satoshi Okuda
Musashi University Lifematics Inc. Waseda University Japan Advanced Institute of
Tokyo, Japan Tokyo, Japan Tokyo, Japan Science and Technology
[email protected] [email protected] [email protected] Nomi, Ishikawa
[email protected]

Nobukazu Yoshioka
Waseda University
Tokyo, Japan
[email protected]

Abstract—In this study, we consider projects for the devel- Under this situation, it is necessary to collect insight from
opment of machine learning (ML) service systems that apply project practices and create reusable knowledge for effectively
ML techniques to enterprise functions, and propose a method undertaking ML projects. In this research, as reusable knowl-
of representing the architecture and design patterns for ML
service systems. Based on the proposed method, we represent edge, we focus on the architecture and design patterns for
the items described in the pattern documents as elements in the ML service systems. Software architecture and design patterns
enterprise architecture modeling and derived a generic model for have been developed mainly for software engineers. In the
ML architecture and designed patterns. By applying the proposed ML projects applied in enterprises, however, as mentioned
method and the generic pattern model, we analyze an existing ML above, members from the business divisions are involved in
design pattern and represent it as a model. Through modeling
practice, we confirm that an effective use scenario occurs when the design of the ML service systems. Therefore, various
using the represented model during the project activities, and project stakeholders should understand the pattern from their
we can revise or enhance the pattern documents consistently by own viewpoints when a pattern for ML service systems is
applying the model. considered in the project. For example, it is important for
Index Terms—Machine learning service system, architecture members from business divisions to understand the objectives
and design pattern, enterprise architecture
of the pattern from a business perspective.
In this study, we propose a method for representing the
I. I NTRODUCTION architecture and design patterns of ML service systems based
on the enterprise architecture (EA) modeling approach. The
Currently, numerous machine learning techniques are avail-
EA modeling is used for representing the new business ser-
able as application programming interfaces (APIs). It is
vices using new IT technologies such that the stakeholders
therefore possible to utilize ML techniques for practical
in business divisions can understand it from their specific
business applications. Considering this situation, enterprises
viewpoints [3]. We try to represent the patterns for ML service
have started implementing such techniques in their business
systems using ArchiMate, which is an EA modeling language,
functions. In this study, we consider projects for developing
and confirm the effectiveness of the EA-based models based
service systems in enterprises using ML APIs. There are new
on an actual example.
features in the service systems using ML techniques (ML
service systems). In addition, when ML techniques are applied The remainder of this paper is structured as follows. In
to these business functions, it is important to acquire training Section II, we describe related studies. In Section III, we
data on the target business domain. This means that sufficient introduce the ML service system and architecture and the
knowledge of, or prior experience in, the target business design patterns for ML service systems. In addition, we
domain is essential. Therefore, the representatives of the IT define the research hypothesis of this research. In Section
division and those of the relevant business division are required IV, we present our EA-based modeling method for the ML
to participate in the project. As results, in ML service system architecture and design patterns, and in Section V, we show
development projects (ML projects), there are many challenges example models for an actual pattern. Finally, the represented
in terms of the requirements, design, implementation, and test example model is discussed in Section VI and the key points
phases [1] [2]. and future studies are summarized in Section VII.

978-1-6654-4488-0/21/$31.00 ©2021 IEEE 245

DOI 10.1109/EDOCW52865.2021.00055
Authorized licensed use limited to: The Technology Library. Downloaded on September 12,2024 at 14:52:33 UTC from IEEE Xplore. Restrictions apply.
 II. R ELATED W ORK is deployed into a runtime ML engine, which then obtains the
input data and provides the output data using this prediction
Through a review of the literature and surveys on some
model.
actual projects, many software engineering challenges to a
project for developing ML systems are identified in [1] [2], In this study, we represent an ML-based service system
and an extraction of knowledge for conducting ML projects is development project carried out by employing an EA modeling
emphasized. approach using ArchiMate [15] as the EA modeling language.
As reusable knowledge for the ML projects, a general The ML-based service system using an EA, as illustrated
workflow for developing ML systems is introduced in [4]. In in Figure 1, was obtained by employing the three business
addition, the role of data scientists in ML system development concepts and three application concepts listed below.
projects is discussed in [5]. By combining these research • Business layer concepts:
outcomes, a project model is proposed that represents the 1) Business service: An explicitly defined and executed
relations between project activities, stakeholders, and project business activity.
goals [6]. In addition, an architecture for representing an entire 2) Business process: A sequence of business activities
system is required as knowledge in practical projects that that produce a planned outcome.
apply big data analytics or machine-learning techniques [7]. 3) Business object: A set of concepts used within a
Several reference architectures for development teams have particular business domain.
also been proposed [8] [9]. In [10], an reference architecture • Application layer concepts:
of the intelligent systems are proposed and this architecture
1) Application service: An explicitly defined and ex-
combines digital strategies and digital architectures [11] with
hibited application behavior.
artificial intelligence.
2) Application component: An element of application
As a knowledge resource for a software system archi-
functionality aligned to the implementation struc-
tecture, the development of architecture and design patterns
ture.
is considered, and many software architecture and design
3) Data object: Data structured for automated process-
patterns have been published. For ML systems, in [12], some
ing.
patterns focusing on the operational stability of ML systems
are introduced. In addition, through a systematic literature
review, software engineering patterns for ML systems are
identified and formalized in [13] [14]. These patterns are
typically described as itemized documents for mainly data
scientists or ML application developers.
EA-modeling approaches are introduced to represent new
business services based on new technologies, allowing various
stakeholders in enterprises to understand the services from
their own viewpoints. In [3], the method for representing a
business model canvas by ArchiMate [15] is proposed, and in
[16], an EA-based innovation modeling is considered.

III. S UBJECT OF R ESEARCH AND R ESEARCH H YPOTHESIS

A. EA and ML-based Service System
It has been stated that there are three types of human Fig. 1. ML-based service system represented by ArchiMate
intelligence: analytical, creative, and practical [17]. In our
research, we considered developing a system with analytical
intelligence for business functions that either support or are B. Architecture design pattern for ML service systems
substitutes for human activities.
For a given input data, analytical intelligence determines A software design pattern is a type of reusable knowledge
and outputs the optimal option from a predefined set [17]. In a in software engineering. In software design patterns, best prac-
workplace, this intelligence is used in routine activities such as tices are formalized such that engineers can use them to solve
replying to service queries or conducting business assessments typical problems that occur when designing an application or
based on customer-provided information. ML techniques are system. Although there is no standard format in many patterns,
used when we replicate this intelligence in a software system. the following items are defined for describing software design
To develop systems using ML techniques, the options for patterns:
the target business domain are first defined; subsequently, the • Intent: Objective of the pattern
example inputs related to each option are collected. An ML • Problem: Forces that the pattern seeks to resolve
model is generated (trained) from a training dataset containing • Solution: Suggested activities to solve the problem
such pairs of options and examples. Subsequently, this model • Context: Environment information on the system

246

Authorized licensed use limited to: The Technology Library. Downloaded on September 12,2024 at 14:52:33 UTC from IEEE Xplore. Restrictions apply.
 • Discussion: Pre-conditions or limitations for applying the notifications are sent to the staff when traffic jams occur. When
pattern a store encounters an adversarial example, e.g., a false positive
For the ML service systems, some security patterns [12] or a false negative, the application sends the corresponding set
and architecture and design patterns [13] [14] are introduced. of images to the cloud for model re-training (the data flows
For example, in [14] a pattern related to the ML system up).
architecture pattern, called a ”data flows up, model flows down Federated learning is a special case of this pattern. It is
with federated learning” pattern, is introduced. Based on the implemented on a Google Keyboard using Android, called
pattern elements introduced above, this model is represented Gboard 2 . When users type texts, their phones store infor-
in Table I. The solution proposed in this pattern is shown in mation about the current context and whether users selected
Figure 2. the suggestions provided by Gboard. Gboard uses an ML
model locally and re-trains the model using this stored data.
TABLE I Thus, Gboard reflects the users’ specific behavior, and the
DATA FLOWS UP, MODEL FLOWS DOWN WITH FEDERATED LEARNING suggestions improve. The difference between the original
PATTERN model and the re-trained models is uploaded to Google Cloud.
!"#$"# !"#$%&'()*'($'+#%,+'()-"'(.%$()*'(-,#/)(0/'$1(2,3()*'(#$'3-4)-%,(
#'$.%$"2,4'(52+'3(%,()*'(6%426(/+'$(%7,(0/'$-'+(2,3(%/)#/)($'+/6)+( At fixed intervals, the base ML model in the cloud is updated
7*-6'(#$'+'$&-,8()*'(/+'$9+(#$-&241:
%&'()$* ;*'(<=(2##6-42)-%,(42,(,%)($')/$,(-)+(#$'3-4)-%,($'+/6)+(-,($'26()-"'(-.(
using these uploaded difference models and re-deployed.
)*'(<=("%3'6(-+(3'#6%1'3()%()*'(46%/3(',&-$%,"',):(;*'(#$'3-4)-%,( The federated learning approach is also implemented in
#'$.%$"2,4'(3'#',3+(%,()*'(/+'$+9(%7,(0/'$-'+:(!.()*'(32)2(4%66'4)'3(
2)()*'(6%426(3'&-4'(-+(+)%$'3(-,)%()*'(46%/3(',&-$%,"',)(.%$()*'( medical imaging systems 3 . To identify diseases from medical
$')$2-,-,8>()*'(/+'$9+(#$-&241(2,3(32)2(4%,.-3',)-26-)1("/+)(5'(
#$'+'$&'3: images using an ML system, huge amounts of medical images
+'),#-'" ? @'#6%1()*'(<=("%3'6()%()*'(6%426(3'&-4': are needed. Such images may not be available in a single
A !,('24*(3'&-4'>()*'(<=("%3'6(-+($'B)$2-,'3(51()*'(6%42661(4%66'4)'3(
32)2: institution, and may not be allowed to be shared owing to
C @-..'$',4'("%3'6+(#$%&-3'3(51(6%426(3'&-4'+(2$'(2&'$28'3(2,3(
/#32)'3(-,)%()*'(<=("%3'6(%,()*'(46%/3: privacy and confidentiality issues. An initial ML model trained
.'"#$/# ! ;*'(2##6-42)-%,(-+($/,,-,8(%,()*'(#'$+%,26(6%426(3'&-4'(+/4*(2+(2( on a small dataset is then deployed to each institution, and
+"2$)(#*%,':
! ;*-+(#$%56'"(+*%/63(5'(4%,+-3'$'3(-,()*'(+1+)'"(3'+-8,(#*2+': the model is re-trained locally using private patient data.
0-12,11-'" ! ;*'$'(2$'(',%/8*(4%"#/)-,8($'+%/$4'+(%,()*'(6%426(3'&-4': Difference models from the institutions are collected and used
! D'B)$2-,-,8(52+'3(%,('24*(/+'$9+(6%8(32)2(3%(,%)(*2&'(2,1(,'82)-&'(
-"#24): to update the ML model.
! E$'3-4)-%,("%3'6+(42,(5'(4%"5-,'3('24*(%)*'$:

C. Research hypothesis
Some architecture and design patterns for ML service sys-
tems provide guidelines to solve multiple problems at once.
For example, by applying the ”data flows up, model flows
down with federated learning” pattern, it is expected that we
can improve the response time for the input query and preserve
the privacy data of the user. For such a pattern, despite the
!"#$%( multiple key elements described in both the problem and intent
fields, we occasionally cannot understand the relationships
!"#$%& !"#$%& between these elements through the freely formatted descrip-
tions. In addition, when developing an ML service system
!"#$%' !"#$%' for enterprise functions, members from both business and IT
divisions are involved from the planning or design phases
during the project. Therefore, various project stakeholders
!"#$%&'#(# )#(#&*+,&,%-(,#././$ should have a common understanding of the patterns applied
0+,&'.**%,%/1%&2+'%34
during the project.
Fig. 2. Structure of the “data flows up, model flows down” pattern For this reason, it is important to represent the architecture
and design patterns for ML service systems as models from
As shown in the context field, the “data flows up, model both business and IT viewpoints. In [18], a method for
flows down” pattern is implemented in the ML system using representing the structure of software design patterns, and in
mobile/edge devices, such as mobile phones, cameras, and [16] a method for representing the business model from both
IoT devices. A major retail company in the USA introduced business and IT viewpoints is considered by using the EA
the ML system to identify traffic problems in their store modeling approach described. In this research, based on such
parking lots 1 . They trained an ML model from the image data studies, we consider a modeling method for the ML design
collected by the parking-lot security cameras and deployed it patterns by using the EA modeling approach. Through the
at each store (model flows down). In the stores, the model proposed method, we represent ML design patterns by using
is applied on real-time images from security cameras, and
2 https://ai.googleblog.com/2017/04/federated-learning-collaborative.html
1 https://www.ibm.com/products/maximo/remote-monitoring 3 https://developer.nvidia.com/gtc/2020/video/s21536-vid

247

Authorized licensed use limited to: The Technology Library. Downloaded on September 12,2024 at 14:52:33 UTC from IEEE Xplore. Restrictions apply.
ArchiMate, which is an EA modeling language. Through a and association, the pattern elements in the ML architecture
modeling example, we confirm the following: and design patterns can be connected to each other.
• Relationships between elements in ML design patterns As a result, as shown in Figure 3, we can represent a generic
can be represented from both business and IT viewpoints. ML architecture and design pattern using ArchiMate.
• Patterns can be described without an over- or under-
description by applying the modeling language.

IV. EA- BASED REPRESENTATION FOR ML ARCHITECTURE

AND DESIGN PATTERNS

A. Subdivision of pattern elements

First, we try to identify common elements described in
each field in the ML design pattern documents. From the
observation of existing ML design patterns, we found the
following elements are described in the problem field.
• Situation to be improved
• Assessment result of the situation to be improved
• Expected goal achieved by applying the pattern

In addition, it was found that the following elements are

typically described in the context field.
• Phase where the pattern is applied
• Device where the system is running
• System user

For these subdivided elements and other existing elements Fig. 3. Generic ML architecture and design pattern represented using
ArchiMate
such as intent, solution, and discussion, we can assign the
model elements defined in ArchiMate. Table II represents the
mapping between pattern elements and EA model elements. V. E XAMPLE
As an example, we apply the proposed method to an ML
TABLE II design pattern called ”encapsulating ML models within rule-
S UBDIVIDED PATTERN ELEMENTS base safeguards,” introduced in [14].
!"##$%&'$($)$&# *+,-./.-$-'0"##$%&'$($)$&# 1($)$&#'.&'2%34.5"#$
Because ML service systems provide outputs based on a
6&#$&# 7,8$3#'9:'0"##$%& 7+#39)$
probability calculation, the system will not always provide
!%9,($) *.#+"#.9&'#9',$'.)0%9/$- ;%./$% correct outputs for various inputs. This means that the ML ser-
2<<$<<)$&#'%$<+(#'9:'#4$' 2<<$<<)$&# vice system may provide a different output for an input whose
<.#+"#.9&'#9',$'.)0%9/$-
=9"('"34.$/$-',>'"00(>.&?'#4$' =9"(
output is pre-determined in advance under the target business-
0"##$%& specific constraints. This situation occasionally occurs when
*9(+#.9& *9(+#.9&'.&'#4$'0"##$%& !%.&3.0($ we introduce ML service systems into the security-critical
@9&#$A# !4"<$'B4$%$'#4$'0"##$%&'.<' C+<.&$<<'0%93$<< business domains. For this situation, this pattern provides a
"00(.$-
solution by introducing a rule-based mechanism and provides
;$/.3$'B4$%$'<><#$)'.<' ;$/.3$
%+&&.&? the rule-matched results as outputs for the specific inputs.
*><#$)'+<$% 23#9% Table III shows the analysis results of the pattern documents
;.<3+<<.9& !%$D39&-.#.9&'9%'(.).#"#.9&'.&' @9&<#%".&#
based on the proposed method.
#4$'0"##$%& From this result, by using the generic representation pattern
model in Figure 3, we can represent the EA-based model
B. Pattern representation by ArchiMate for this ML design pattern, as shown in Figure 4. Through
this representation, we can understand this pattern from the
As described above, in the pattern documents, relationships
following perspectives:
between key elements are occasionally not clearly described,
which hinders the common understanding on the patterns • How can we benefit by applying the pattern?

among stakeholders. In the EA modeling using ArchiMate, we • When or where will the pattern be applied during the

can represent the relationships between elements. For example, development project?
the solution, represented as a principle, is considered to realize • What types of ML service systems are considered in this

the object of the pattern, which is represented as an outcome. pattern?

This relationship can be represented using the realization In ML service system development projects, by looking at
notation defined in ArchiMate. By using relationships such the upper items of Figure4, described through the motivation
as realization, composition, access, flow, serving, assignment, extension in ArchiMate, project members from the business

248

Authorized licensed use limited to: The Technology Library. Downloaded on September 12,2024 at 14:52:33 UTC from IEEE Xplore. Restrictions apply.
 TABLE III
ML ARCHITECTURE AND DESIGN PATTERN ELEMENTS ( ENCAPSULATING ML MODELS WITHIN RULE - BASED SAFEGUARDS )

!"##$%&'$($)$&# *+,-./.-$-'0"##$%&'$($)$&# 1($)$&#'.&'#2$'$&3"04+("#$'56')7-$(4'8.#2.&'%+($9,"4$'4":$;+"%-4 0"##$%&

<&#$&# =,>$3#'7:'0"##$%& ?$"(.&;'8.#2'#2$'+&3$%#".&#@'"&-'37)0($A.#@'7:'#2$'7+#0+#'+4.&;'56')7-$(4

!%7,($) *.#+"#.7&'#7',$'.)0%7/$- *":$#@

B44$44)$&#'%$4+(#'7:'#2$' ! *.&3$'0%$-.3#.7&',@'56')7-$(4'.4'7:#$&'-.::.3+(#'#7';+"%"&#$$'37%%$3#&$44C'56'
4.#+"#.7&'#7',$'.)0%7/$- )7-$(4'3"&&7#',$'+4$-'-.%$3#(@':7%'4":$#@9%$("#$-':+&3#.7&4D
! 56')7-$(4'"%$'+&4#",($'"&-'/+(&$%",($'#7'"-/$%4"%."('"##"3E4C'-"#"'&7.4$C'"&-'-%.:#D
F7"('"32.$/$-',@'"00(@.&;' G$-+3.&;'4":$#@'%.4E4
#2$'0"##$%&
*7(+#.7& *+;;$4#$-'"3#./.#.$4'#7' 1&3"04+("#$'56')7-$(4'.&'"'-$#$%).&.4#.3'"&-'/$%.:.",($'%+($9,"4$-')$32"&.4)D
47(/$'#2$'0%7,($)
H7&#$A# !2"4$'82$%$'#2$'0"##$%&'.4' =/$%"(('4@4#$)'"&-'47:#8"%$'-$4.;&'02"4$C'.&3(+-.&;'56')7-$(4
"00(.$-
?$/.3$'82$%$'4@4#$)'.4' ?$/.3$4'82$%$'#2$'4":$#@9%$("#$-'4@4#$)4'"%$'%+&&.&;
%+&&.&;
*@4#$)'+4$% I4$%4'7:'#2$'4":$#@9%$("#$-'4@4#$)4D

?.43+44.7& !%$937&-.#.7&'7%'(.).#"#.7&' JB
.&'#2$'0"##$%&

Fig. 4. ML architecture and design pattern represented using ArchiMate (encapsulating ML models within rule-based safeguards)

division can understand why the pattern will be applied in the tion in the pattern document. Therefore, it is expected that we
project. In addition, if some pattern candidates exist for solving can consistently revise or enhance the pattern documents by
a specific problem, by focusing on the lower items described updating the EA-based ML design pattern model proposed in
by the application or technology elements of the ArchiMate this research.
and the actor, development teams can assess which pattern
should be selected for the use case considered in the project. VI. D ISCUSSION
Through the practices applied to the project, various types Through the representation of an existing ML design pattern
of insights are collected. As a result, ML design patters may using the proposed method, We can expect that the pattern
be revised or enhanced based on such insights. Each model elements and the relationships between elements can be rep-
element in Figure4 is aligned with the corresponding descrip- resented from both business and IT viewpoints, and project

249

Authorized licensed use limited to: The Technology Library. Downloaded on September 12,2024 at 14:52:33 UTC from IEEE Xplore. Restrictions apply.
stakeholders can have a common understanding of the pattern. R EFERENCES
In addition, it is expected that the pattern document can [1] F. Kumeno, “Software engineering challenges for machine learning
be revised or consistently enhanced by updating the pattern applications: A literature review,” Intelligent Decision Technologies,
model. Further investigation through the practices of ML vol. 13, pp. 463 – 476, 2019.
[2] L. E. Lwakatare, A. Raj, J. Bosch, H. H. Olsson, and I. Crnkovic, “A tax-
projects in various domains will be needed for validating these onomy of software engineering challenges for machine learning systems:
items. This remains as our future work. An empirical investigation,” in Proceedings of the 20th International
Conference on Agile Software Development (XP), 2019, pp. 227 – 243.
The design patterns are not independent of each other and [3] L. O. Meertens, M. E. Jacob, L. J. M. Nieuwenhuis, M. J. V. Sinderen,
many are closely related either directly or indirectly. For H. Jonkers, and D. Quartel, “Mapping the business model canvas to
example, in [18], design patters are grouped according to their archimate,” in Proceedings of the 27th Annual ACM Symposium Applied
Computing (SAC), 2012, pp. 1694 – 1702.
quality. This is the same for ML design patterns. In [19], [4] S. Amershi, A. Begel, C. Bird, R. Deliner, H. Gall, E. Kamar, N. N. B.
ML design patterns are categorized into six groups, including Nushi, and T. Zimmermann, “Software engineering for machine learn-
the data representation, reproducibility, and resilience, and the ing: A case study,” in Proceedings of the 41st International Conference
on Software Engineering, 2019, pp. 291 – 300.
relationships between patterns are summarized as a network [5] M. Kim, T. Zimmermann, R. DeLine, and A. Begel, “The emerging
diagram. Using an EA-based representation of the ML design role of data scientists on software development teams,” in Proceedings
patterns, we can expect that the patterns can be related to each of the 38th International Conference on Software Engineering, 2016,
pp. 96–107.
other from various viewpoints such as the goal, assessment, [6] H. Takeuchi and S. Yamamoto, “AI service system development using
actor, and device. Extending the proposed method such that it enterprise architecture modeling,” in Proceedings of the 23rd Interna-
can represent the relationships between ML design patterns is tional Conference on Knowledge-Based and Intelligent Information &
Engineering Systems(Procedia Computer Science vol. 159), 2019, pp.
also reserved for future study. 923 – 932.
[7] S. Earley, “Analytics, machine learning, and the internet of things,” IEEE
VII. C ONCLUSION ITPro, vol. 17, no. 1, pp. 10–13, 2015.
[8] Y. Demchenko, C. de Last, and P. Membrey, “Defining architecture com-
ponents of the big data ecosystem,” in Proceedings of the International
In this study, we focused on projects for the development Conference on Collaboration Technologies and Systems (CTS), 2014,
of ML service systems that apply ML techniques to enterprise pp. 104–112.
functions. We considered a method of representing the design [9] J. Heit, J. Liu, and M. Shah, “An architecture for the deployment
of statistical models for the big data era,” in Proceedings of IEEE
patterns for ML service systems based on the EA-modeling International Conference on Big Data, 2016, pp. 1377–1384.
approach. Using the proposed method, we subdivided the [10] A. Zimmermann, R. Schmidt, D. Jugel, and M. Möhring, “Evolution
items conventionally applied in the pattern documents, and of enterprise architecture for intelligent digital systems,” in Proceed-
ings of the 14th International Conference on Research Challenges on
to each subdivided item we assigned the EA model element Information Science, 2020, pp. 145 – 153.
defined in ArchiMate, an EA modeling language. We also [11] A. Zimmermann, R. Schmidt, K. Sandkuhl, D. Jugel, J. Bogner, and
defined the relationships between the subdivided items and M. Möhring, “Evolution of enterprise architecture for digital transfor-
mation,” in Proceedings of the IEEE 22nd International Enterprise
connected the EA model elements. As a result, we derived a Distributed Object Computing Workshop, 2018, pp. 87 – 96.
generic ML architecture and design pattern model represented [12] H. Yokoyama, “Machine learning system architectural pattern for im-
by ArchiMate. proving operational stability,” in Proceedings of IEEE International
Conference on Software Architecture Companion, 2019, pp. 267 – 274.
Based on the proposed method and the generic pattern [13] H. Washizaki, H. Uchida, F. Khomh, and Y.-G. Guéhéneuc, “Software
model, we analyzed an ML design pattern called ”encapsulat- engineering patterns for machine learning applications (sep4mla),” in
Proceedings of the 9th Asian Conference on Pattern Languages of
ing ML models within rule-base safeguards” and represented Programs (AsianPLoP 2020), 2020.
it as an EA-based model. Through the represented model, [14] H. Washizaki, F. Khomh, Y.-G. Guéhéneuc, H. Takeuchi, S. Okuda,
we confirmed that project members from both business and N. Natori, and N. Shioura, “Software engineering patterns for machine
learning applications (sep4mla) - part 2,” in Proceedings of the 27th
IT divisions can be expected to understand the ML design Conference on Pattern Languages of Programs (PLoP 2020), 2020.
pattern applied in the project from their own perspectives. In [15] The Open Group, ArchiMate 3.1 – A Pocket Guide. Van Hares
addition, we found that we can revise or enhance the ML Publishing, 2019.
[16] S. Yamamoto, N. I. Olayan, and J. Fujieda, “e-healthcare service design
design pattern documents consistently by updating the model using model based jobs theory,” in Proceedings of International Confer-
diagram. As a future work, validating the proposed method and ence on Intelligent Interactive Multimedia Systems and Services(KES-
a generic model through practices applied in various domains IIMSS), 2018, pp. 198–207.
[17] R. J. Sternberg, Successful Intelligence: How Practical and Creative
and representing the relationships between ML design patterns Intelligence Determines Success in Life. Simon & Schuster, 1996.
based on the proposed generic model will be considered. [18] N.-L. Hsueh and W.-H. Shen, “Handling nonfunctional and conflicting
requirements with design patterns,” in Proceedings of the 11th Asia-
Pacific Software Engineering Conference (APSEC 04), 2004, pp. 608
ACKNOWLEDGMENT – 615.
[19] V. Lakshmanan, S. Robinson, and M. Mann, Machine Learning Design
This work was supported by a JSPS Grant-in-Aid for Sci- Patterns: Solutions to Common Challenges in Data Preparation, Model
entific Research (KAKENHI), Grant No. JP19K20416; JST- Building, and MLOps. O’Reilly, 2020.
Mirai Project (Engineerable AI Techniques for Practical Ap-
plications of High-Quality Machine Learning-based Systems),
Grant No. JPMJMI20B8; and ROIS NII Open Collaborative
Research 2021-21S0801.

250

Authorized licensed use limited to: The Technology Library. Downloaded on September 12,2024 at 14:52:33 UTC from IEEE Xplore. Restrictions apply.