A Quality-driven Machine Learning Governance Architecture
for Self-adaptive Edge Clouds
Claus Pahl, Shelernaz Azimi, Hamid R. Barzegar and Nabil El Ioini
Free University of Bozen-Bolzano, Bolzano, Italy
Machine Learning, Quality, Edge Cloud, Cloud Controller, Resource Management, AI Engineering, AI
Governance, Explainable AI.
Self-adaptive systems such as clouds and edge clouds are more and more using Machine Learning (ML)
techniques if sufficient data is available to create respective ML models. Self-adaptive systems are built around
a controller that, based on monitored system data as input, generate actions to maintain the system in question
within expected quality ranges. Machine learning (ML) can help to create controllers for self-adaptive systems
such as edge clouds. However, because ML-created controllers are created without a direct full control by
expert software developers, quality needs to be specifically looked at, requiring a better understanding of the
ML models. Here, we explore a quality-oriented management and governance architecture for self-adaptive
edge controllers. The concrete objective here is the validation of a reference governance architecture for edge
cloud systems that facilitates ML controller quality management in a feedback loop.
Edge cloud systems are more and more using Ma-
chine Learning (ML) techniques if sufficient data is
available to create ML models. A typical example is
a self-adaptive system built around a controller that,
based on monitored system data as the input, gener-
ates actions to maintain the system in question within
expected quality ranges. Self-adaptive systems are
prevalent in many IoT and edge cloud environments
where manual adjustment is not feasible or not reli-
able. Virtualized systems are ideally suitable to apply
some sort of control-theoretic mechanism to continu-
ously and automatically adjust a dynamic system.
Our focus is edge clouds, i.e., connected device
infrastructures that built on virtualization to allow dy-
namic resource management in line with application
demands that create elasticity in the responses to de-
mand changes. The management of the resources in
terms of topology configuration, resource utilization
and application is nowadays often based on models
that are machine learned.
Machine learning can thus help to create con-
trollers for self-adaptive edge cloud systems where
automation is of critical importance (Mendonca et al.,
2021). However, due to the ML approach to con-
troller creation without a direct full control by expert
software developers, quality needs to be specifically
looked at. In broader terms, what is needed is a soft-
ware engineering approach for ML-constructed soft-
ware that addresses the specific quality concerns, of-
ten referred to as AI Engineering (Lwakatare et al.,
2019). AI Governance based on better understand-
ing (often referred to as explainability) and trans-
parency of how the ML model works is here central in
this wider AI engineering objective (Pahl and Azimi,
The observations above on ML-generated con-
trollers and quality concerns related to this type of
software (Azimi and Pahl, 2020b; Azimi and Pahl,
2020a; Azimi and Pahl, 2021) lead to challenges
that we aim to address in our chosen technical ap-
plication domain of edge cloud architectures (Pahl
et al., 2018; Taibi et al., 2020; El Ioini et al., 2021).
We develop a quality-driven governance architecture
for self-adaptive edge-cloud applications that dynam-
ically adjusts these their controllers through continu-
ous model creation and improvements. The novelty
lies in a comprehensive quality improvement frame-
work beyond effectiveness of the controller only.
While both edge cloud management and in partic-
ular ML quality and governance are broad and chal-
lenging concerns, we will limit our investigation here
to a clearly defined scope: resource-constrained edge
clusters as the architectural setting and reinforcement
learning as the ML technique for the controller itself
Pahl, C., Azimi, S., Barzegar, H. and El Ioini, N.
A Quality-driven Machine Learning Governance Architecture for Self-adaptive Edge Clouds.
DOI: 10.5220/0011107000003200
In Proceedings of the 12th International Conference on Cloud Computing and Services Science (CLOSER 2022), pages 305-312
ISBN: 978-989-758-570-8; ISSN: 2184-5042
2022 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
to determine self-adaptation mechanisms. The tech-
nical challenges to be addressed here include, firstly,
scoring functions for considered qualities, including
computability (for self-adaptable) and effective usage
as the primary necessary aim and explainability as a
second critical aim and, secondly, remedial strategies
to address any detected quality deficiencies.
The paper is structured as follows. We first in-
troduce a conceptual framework in Section 2, before
analysing the state-of-the-art in Section 3. Section 4
presents our proposed architecture, which is the dis-
cussed in Section 5. We conclude with a summary
and future work in Section 6.
2.1 Use Case and Motivation
In order to achieve meaningful results within the con-
straints given, we focus on a specific architectural set-
ting: IoT and edge cloud architectures. Furthermore,
the evaluation will involve concrete use cases from
the domain of smart traffic and smart mobility.
The concrete setting here is edge cloud resource
management (Scolati et al., 2019; von Leon et al.,
2019). System adaption is required for edge cloud re-
source configuration, involving: monitor resource uti-
lization and application performance and apply rules
(an ML-generated rule model) for resource adapta-
tion to meet performance and cost requirements. The
rules adapt the resource configuration (e.g., size) to
improve the performance/cost ratio. The chosen ML
techniques here is reinforcement learning that works
on a reward principle applicable in the self-adaptation
loop. Finally, enact configuration change recommen-
dation. The problem shall be described using a con-
crete use case:
Problem: A resource controller for edge adap-
tation might suggests: if Workload > 80% then
Double(resource size). The question is whether
this recommendation is correct and overall the
recommendation space is complete.
Solution: Here a generated ML model could pro-
vide a recommendation for a 60% workload as a
verified test case. Then, this recommendation can
be scaled up to 80%.
ML-driven controller (model) creation and automated
dynamic adaptation (the adaption by the controller
is based on the rule model) cannot be done without
proper quality monitoring. For this, the model quality
needs to be assessed through score function regard-
ing effectiveness, e.g., based on high accuracy. De-
tected quality deficiencies need to be subjected to a
root cause analysis and suitable recommendation and
enactment of remedies need to be determined.
2.2 Conceptual Challenges
Quality management of ML models in generated soft-
ware and its link to underlying raw/training data is at
the core of the problem here. Ground truth, which
is the accuracy against the real world, in our cases
means whether the system performance actually im-
proves, if up-scaling was recommended.
Remedies include to improve data labelling, e.g.,
automate critical situation assessment (high risk of
failure, based on past experience), which have a prob-
ability of discrimination and could be biased, be that
through pre-processing (before ML training) and in-
processing (while training). Research tasks are to
find bias and remove bias (through a control loop),
e.g., using favourable data labels. Examples are
could smaller or bigger device clusters be favoured
wrongly or specific types of recommended topologies
or recommended configuration sizes (messages, stor-
age etc)? Furthermore, explainability of the controller
recommendations is a critical ingredient to map ob-
served ML model deficiencies back to system-level
properties (via the monitored input data to the ML
model creation).
Our focus will be on the controller environment,
i.e., the sensing devices that monitor the performance
of the system and the actuators that enforce remedies
proposed by the controller. Thus, we will consider
faultiness in this environment as the root causes of
quality problems. These faults could include sensor
faults or communication faults in the network. In dy-
namic systems, a further complication is the need for
continuous evaluation due to changing circumstances.
We review related work, specifically focusing on ML
for dynamic edge and IoT settings, before analysing
the limitations and challenges.
3.1 Related Work
The application of ML techniques ranges from non-
critical business applications to autonomous safety-
critical software (Karkouch et al., 2016). Specifi-
cally, in distributed settings, the need for software de-
pendability rises. A survey regarding the impact of
ML on software development (Tokunaga et al., 2016)
concludes that software practitioners still struggle to
CLOSER 2022 - 12th International Conference on Cloud Computing and Services Science
operationalize and standardize the software develop-
ment practices of systems using ML. Instead of func-
tional requirements that are often the focus in non-
ML software, quantitative non-functional factors such
as accuracy form central requirements for business-
critical or safety-critical ML systems, which are in
some business-critical or safety-critical domains of
great importance.
It is noted that these systems are highly coupled.
For instance, the performance of models is dependent
on the quality of data processing. Poor data process-
ing could prevent an effective ML model construc-
tion. In ML systems, “too low” and “too high” scores
for performance measures as testing results both in-
dicate defects. Not always is pre-construction valida-
tion [29] possible; thus, we aim here at an a-posteriori
analysis to remedy problems.
The quality of raw data, the machine learning
process perspective and the machine learning model
quality are key building blocks in the construction
process. Raw or source data quality has been in-
vestigated (Mohammedameen et al., 2019), result-
ing in quality frameworks that our earlier selection of
quality attributes is based on. In (Mohammedameen
et al., 2019)[23], data quality problems where clas-
sified into two groups of context-independent and
context-dependent from the data and user perspective.
In (De Hoog et al., 2019), a new architecture based
on Blockchain technology was proposed to improve
data quality and false data detection. In (Sridhar et al.,
2018), a prototype of a distributed architecture for IoT
was also presented, providing supporting algorithms
for the assessment of data quality and security.
The ML process perspective is discussed in
(Amershi et al., 2019). A machine learning workflow
with nine stages is presented in which the early stages
are data oriented. Usually the workflows connected
to machine learning are non-linear and often contain
feedback loops to previous stages in order to rem-
edy quality concerns. If the system contains multi-
ple, interconnected ML components, quality becomes
even more critical. Investigating a broader loop from
the final ML function construction stages to the ini-
tial data and ML training configuration stages has not
been comprehensively attempted yet.
Another aspect is the machine learning model
layer. Different supervised learning approaches were
used. Specific quality metrics apply to ML tech-
niques. The area under the receiver operating charac-
teristic curve (AUC) is an example of quality for clas-
sification models. In (Sicari et al., 2016), a solution
for model governance in production machine learn-
ing can build on provenance information to trace the
origin of an ML prediction solution in order to iden-
tify the root cause of an observed problem symptom.
Also the quality of data in ML has been investigated.
An application use case was presented, but without a
systematic coverage of quality aspects. Data quality
as the root cause is important in many ML-supported
applications. In (Deja, 2019), the authors inves-
tigate high-energy physics experiments. The work
presented in (Deja, 2019) serves as an IoT setting.
Some works highlight the need for a systematic, auto-
mated approach to achieve higher accuracy to remedy
training problems arising from manual data labelling
(Sheng et al., 2008). Here, ML techniques such as iso-
lation forests as classification techniques or autoen-
coders as neural networks are looked at.
While most previous work looked for root causes
of quality deficiencies in the ML construction, the IoT
and edge cloud environment also needs to be consid-
ered as a consequence of the uncertainty of sensory
data as a problem cause (Efron, 2020). Thus data
quality needs a more prominent role in the construc-
tion process. (Efron, 2020) covers IoT root causes in
the analysis, but not training/ML data problems.
3.2 Analysis and Challenges
Our task is to condense the different individual quality
concerns into an integrated quality model that takes
on board lessons learned from (De Hoog et al., 2019;
Ehrlinger et al., 2019; Fang et al., 2016), but provides
a closed feedback loop. We apply our proposed ar-
chitecture to resource management and orchestration
in edge clouds (Hong and Varghese, 2019), where
controllers manage systems self-adaptively. Here, the
compute, storage or network resources can be con-
figured dynamically through a self-adaptation mech-
anism. Network configurations as one of the plat-
form concerns have been investigated by a number of
authors (Mantri et al., 2013; Tokunaga et al., 2016;
Femminella and Reali, 2019). Another strategy is
the dynamic allocation and management of tasks in
distributed environments. Workload balancing has
been a proposed solution (Zhao et al., 2019). Some
authors have investigated specific applications such
as streaming and other data-intensive contexts (Mo-
hammedameen et al., 2019), where the interplay of
communication, computation and storage needs to be
coordinated. Equally, specific application domains
such as vehicular networks have been looked at (Javed
et al., 2020).
Our direction here is a specific type of architec-
ture with layered and clustered edge clouds that are
for instance suitable for automotive applications and
networks. This requires a combination of horizon-
tal and vertical orchestration techniques that have not
A Quality-driven Machine Learning Governance Architecture for Self-adaptive Edge Clouds
been sufficiently explored for low-latency and high
data volume applications.
Machine learning has been used in some of these
advanced architectures (Wang et al., 2020). For in-
stance, reinforcement learning (RL) is a promising so-
lution. We also use reinforcement learning and build
on our experience in QL and SARSA as two RL ap-
proaches that we used for resource management in
central clouds and adapt this to edge architectures
(Jamshidi et al., 2015; Arabnejad et al., 2017).
In the previous section, we already identified some
concrete challenges and outlined the key architectural
principles. This shall now be used to define a quality-
driven architectural governance framework for edge
cloud resource management. At the core is a ML-
constructed controller.
4.1 ML-generated Controllers and
Scoring Functions
Controllers that are ML-generated create the chal-
lenge of assessing and maintaining quality in a largely
automated process. Intelligent quality management
embedded into a systematic engineering framework
is needed. The controller solution for self-adaptive
system is build on reinforcement learning (RL) tech-
niques that we have already successfully used for
centralised cloud architectures (Jamshidi et al., 2015;
Arabnejad et al., 2017), here adapted to the distributed
and constrained edge context. It aims to maximize
the notion of a cumulative reward. RL does not as-
sume knowledge of an exact mathematical model and
is thus suitable where exact methods become infeasi-
ble due to resource constraints and dynamic changes.
The controller manages workload and perfor-
mance as qualities of the system in question, but
the controller needs to be assessed in terms of ML
qualities. Non-functional properties and respective
scoring functions are relevant here for the controller
(rather than functional testing), but this requires tar-
geted score functions for quality measurement of the
controller: accuracy of the controller actions is here
the central scoring function, i.e., how accurate are
the analyses of the controller and the remedial actions
taken. We need to define score functions that are im-
mediately computable (e.g., via group cardinalities as
a suggestion), rather than observing future behaviour,
which cause often unacceptable time delays.
4.2 A Reference Architecture for Edge
Cloud Controller Quality
Adaptors and decision models at the core of con-
trollers are suitable for ML-based creation. The out-
come of the ML model is to implement an action, e.g.,
to adapt resources, divert traffic (IoT, roads), or to in-
struct machines. This should be a dynamic control
loop, guided by defined quality goals. We use IoT-
traffic use cases for the experimental validation.
4.2.1 Reference Governance Architecture
Our technical objectives are organised as follows into
two layers to be reflected in our architectural frame-
work: We develop a controller for self-adaptive edge
cloud systems based on reinforcement learning as the
ML technique. We develop an intelligent quality
monitoring and analysis framework aligned with the
needs of ML-generated controller software for self-
adaptive software. This architecture aims to support
quality management for ML-based adaptors. Figure
1 builds on a MAPE-K loop (Monitoring, Analysing,
Planning, Executing based on Knowledge) in two
layers. The upper loop is the focus, but needs to take
into account lower layer behaviour.
Figure 1: System Architecture a lower edge controller
loop and an upper ML quality management loop.
4.2.2 ML Controller and Quality Management
Important questions concern the full automation of
the upper loop. The full automation is, however, not
the objective here in this investigation. Nonetheless,
this ultimate objective shall guide the work and shall
aid its solution in the future. In concrete terms, the
challenges are: Automate testing (M in MAPE), Au-
tomate the test evaluation (A in MAPE), Device a
ML learning adjustment (P in MAPE). For this work,
however, the upper loop is not meant be fully self-
adaptive. Our objective here is to develop a core un-
derstanding that would allow full automation in the
CLOSER 2022 - 12th International Conference on Cloud Computing and Services Science
future. Thus, the objective here is to investigate the
validity of the governance architecture, i.e., to demon-
strate that an automation is beneficial and feasible.
For the MAPE-K controller design, we deploy a
MAPE-K architecture pattern for the adaption as a
specific type of an ML model self-adaption regarding
accuracy. The evaluation of the controller is based
on testing by checking variations and their effect the
users experience regarding some metrics. The meta-
layer upper MAPE-K loop is designed as follows:
M: to consider are score functions for ML model
quality, i.e., adaptor quality based on accuracy
that is if possible linked back to application data
quality, e.g., cloud resource utilization, applica-
tion performance, with correctness and complete-
ness as ML quality concerns.
A: the analysis consisting of a root cause analy-
sis for ML model quality problems and feeding
into explainability concerns through partial de-
pendency determination in order to identify what
system factors can improve target quality most.
P/E: To recommend and execute these recommen-
dations, i.e., a rule update for the cloud adaptor
with ML model creation configuration: training
data size/ratio — linked to raw data correction (at
least a recommendation).
This implements a meta-learning process that is
not only learning to adapt the system, but learn-
ing to adapt the adaptor through a continuous
testing/experimentation process that implements a
knowledge learning layer. Continuous evaluation is,
as in any self-adaptive system, a key component. The
ultimate aim is to automate the model variation by
MAPE-ing the ML model construction (meta-level
optimization), e.g., labelling or test size/ratio.
The target infrastructure is an IoT and edge cloud en-
vironment, where small devices (IoT) at the edge are
linked to central cloud data centres.
Mobile edge use case: The key evaluation sce-
nario is a video streaming application. The respective
infrastructure will be fully developed as part of the EU
H2020 project 5G-CARMEN that we are a partner of
(Pomalo et al., 2020; Pomalo et al., 2020; Barzegar
et al., 2020; Barzegar et al., 2021).
Lightweight edge devices and containers: We use
a cluster of lightweight edge devices (specifically
Raspberry Pi devices), deploying container-based ap-
plications on these.
Cloud and Edge Computing Lab (CECL): The
CECL enables virtualization, cloud and edge com-
puting experimentation. It consists of 3 HP DL380
gen10 servers (6 Virtual CPUs, 192GB RAM, 6 TB
SSD) as central cloud/edge servers and 80 Raspberry
Pi IV (BCM2711, Cortex-A72 quad-core (ARMv8)
64-bit SoC @ 1.5GHz, 4G RAM, 16G SD card) as
lightweight edge devices.
Currently, a prototype is under development. A
first controller version running on an RPi cluster is
available (Gand et al., 2020).
5.1 Validation
We discuss the objectives, the setup of the validation
experiments and report on our observations.
5.1.1 Evaluation Objectives
Results of the controller implementation – presenting
the evaluation of the lower layer of our architecture
from Fig. 1 have been presented in (Gand et al.,
2020). Here we focus on results for the upper layer
of Fig. 1. In general, the evaluation of a quality con-
troller (the upper layer) needs to cover the follow-
ing solution components and respective strategies to
evaluate: Regarding the monitoring of the controller
quality, we need to carry out anomaly detection in the
CECL lab. Evaluation activities are to induce anoma-
lies into the testbed based on realistic patterns and to
measure the impact by determining the accuracy as
the selected scoring function of induced fault patterns.
In this paper, the primary objective is to evaluate
the usefulness of the reference governance architec-
ture. While the lower layer is already well explored,
whether the proposed upper layer quality controller
is beneficial and feasible remains an open question.
For this, we need to demonstrate that by monitor-
ing the controller, the root cause of detected quality
anomalies can be detected and remedial actions be
proposed, following the MAPE-K loop: M: monitor
accuracy (and also related precision and recall) of the
ML model. A: analyse the anomalies in the controller
quality over time. P: plan a response using a rule sys-
tem for remedial actions. E: provide a concrete, ac-
tionable remedy recommendation.
5.1.2 Experiments
In order to answer this research question, we con-
ducted simulation experiments by inducing anomalies
into two different input data sets: Data set 1 is a snap-
shot data of 72 rows counting volume of utilisation,
14 features that are largely independent. Data set 2 is
A Quality-driven Machine Learning Governance Architecture for Self-adaptive Edge Clouds
a time series data of 197 rows collected at 49 data col-
lection points on environmental data, 15 features that
are partially interdependent. These were then anal-
ysed in terms of their impact of ML model accuracy
(as well as precision and recall) and if a root cause and
a respective remedy could be determined in order to
complete the MAPE-loop described above. The use
case will be the scenario already discussed: a smart
traffic/mobility scenario.
The anomalies that we induced into the two data
sets were:
incompleteness: to simulate that monitors do not
provide any data or that the connections between
devices is down.
incorrectness: to simulate that monitors provide
incorrect data (because of faultiness of the moni-
tors themselves or transmission faults).
The anomalies were systematically created, covering
the following dimensions:
extend or degree of incompleteness or incorrect-
ness: We analysed different degrees of incom-
pleteness and used incorrect data ranging from
slightly out of normal ranges up to extreme and
impossible values.
variability of anomalies: the two types were both
induced in a random ways as well as clustered on
both data rows and features.
The example in Fig. 2 that shows blocks (clusters) of
incorrect data for various features here using implau-
sible negative values that could in an edge cloud set-
ting indicate transmission or conversion errors.
These anomalies allow to map the anomaly types
to systems faults, for instance. clustered incomplete-
ness of rows can be associated with local network
faults or time-clustered incorrectness can be associ-
ated with sensor faults. We created time series of
anomalous model quality. These times series were
then analysed with regard to possible change patterns
regarding accuracy, precision and recall.
5.1.3 Results
We observed accuracy, precision and recall for the
ML-generated model. Fig. 3 shows an example of a
specific test case for increasing incorrect sensor val-
ues. Overall, more than 50 individual settings were
explored. These graphs were then compared in terms
of gradient, function type, and other criteria. Some
key observations from the analysis of the ML model
accuracy, precision and recall graphs are as follows:
Incorrectness is more significant than incomplete-
ness. The incorrectness has a bigger effect on the
accuracy than the incompleteness. A possible rea-
son here is that in incompleteness the machine
learning an ML tool may ignore missing data rows
or features and not include these in the predic-
tions and calculations. However, for incorrectness
a tool is forced to use all values, be them correct
or not. Thus, it cannot control or minimize the
negative impact on accuracy.
Rows in a data set are more significant than fea-
tures. Missing or invalid rows have a stronger im-
pact on the accuracy than missing or invalid fea-
tures in the data set tables. Here a number of are
possible. A probable one is the fact that dealing
with a complete missing or invalid row is more
difficult than dealing with some missing or invalid
features. Resolving the reduction of accuracy is
consequently more difficult with missing/invalid
rows than with missing/invalid features, resp.
These observations represent identifiable change pat-
terns, e.g., increasing incorrectness having a more sig-
nificant (negative) impact on accuracy). As already
mentioned, these changes can then be associated to
root causes, e.g.,
significant accuracy changes point to incorrect-
ness cause most likely by sensor faults,
faultiness of individual sensors has less impact
then communication faults.
Using this kind of rule system, useful recommenda-
tions for remedial actions such as checking or replac-
ing faulty sensors can be given. Thus, the experiments
demonstrate that a quality controller is feasible and
the reference architecture is valid.
Although our focus here is on edge cloud controllers
as a typical example where automation is necessary
and beneficial, the problem of quality control for ML-
generated software is a broader one. Controllers can
be effectively generated using machine learning from
past monitoring and reaction data collected during
dynamic resource management. While the benefits
of automation and dynamic adaptation of application
systems is clear, more work needs to be done on on-
going quality management for the ML-generated con-
trollers themselves, using a meta-level feedback loop
on top of the application management one. Managing
AI (or more specifically ML) quality is a concern of
AI governance and AI engineering that put more em-
phasis on understanding the AI systems in question
and making quality management a more systematic
engineering approach, respectively.
CLOSER 2022 - 12th International Conference on Cloud Computing and Services Science
Figure 2: Sample anomaly-induced data set: clustered blocks of invalid features (implausible values).
Figure 3: Observation of ML quality metrics (accuracy) for
different incorrect (different degrees of incorrectness show-
ing more impact for more unexpected/implausible values).
For our edge cloud setting, we have proposed a
reference governance architecture here that embodies
the two feedback loops in a layered setting. A com-
mon problem that applies here is that the quality of the
controller is observable to users, but the root causes
that might negatively impact on that quality are not
under the control of the user or are not easily visible.
Based on experiments in a simulated setting, we
have demonstrated that the proposed governance ar-
chitecture is feasible, i.e., the controller quality can
be monitored and an analysis of usually hidden root
causes of root causes is possible, also allowing bene-
ficial remedial actions to be proposed.
This paper reports on the first stage of the over-
all ambition, i.e., the validity of the reference model
was demonstrated through experiments. In a second
stage as part of the future work, we aim to fully im-
plemented the layered feedback loop and validate the
initial results with concrete implementations and real
anomalous data from that system. Nonetheless, since
the simulated experiments conducted here are based
on induced anomalies that cover all common fault and
anomaly situations in this architectural setting, we can
be confident about the validity of our results.
Amershi, S., Begel, A., Bird, C., DeLine, R., Gall, H., Ka-
mar, E., Nagappan, N., Nushi, B., and Zimmermann,
T. (2019). Software engineering for machine learning:
A case study. In Intl Conf on Software Engineering -
Software Engineering in Practice track. IEEE.
Arabnejad, H., Pahl, C., Jamshidi, P., and Estrada, G.
(2017). A comparison of reinforcement learning tech-
niques for fuzzy cloud auto-scaling. In International
Symposium on Cluster, Cloud and Grid Computing,
CCGRID 2017. IEEE Computer Society / ACM.
Azimi, S. and Pahl, C. (2020a). A layered quality frame-
work in machine learning driven data and information
models. In 22nd International Conference on Enter-
prise Information Systems.
Azimi, S. and Pahl, C. (2020b). Particle swarm optimization
for performance management in multi-cluster iot edge
architectures. In International Conference on Cloud
Computing and Services Science CLOSER.
Azimi, S. and Pahl, C. (2021). The effect of iot data
completeness and correctness on explainable machine
learning models. In Database and Expert Systems Ap-
plications. Springer.
Barzegar, H. R., Ioini, N. E., Le, V. T., and Pahl, C. (2021).
5g-carmen: Service continuity in 5g-enabled edge
clouds. In Advances in Service-Oriented and Cloud
Computing. Springer.
Barzegar, H. R., Le, V. T., Ioini, N. E., and Pahl, C. (2020).
Service continuity for ccam platform in 5g-carmen.
In 2020 International Wireless Communications and
Mobile Computing (IWCMC), pages 1764–1769.
A Quality-driven Machine Learning Governance Architecture for Self-adaptive Edge Clouds
De Hoog, J., Mercelis, S., and Hellinckx, P. (2019). Improv-
ing machine learning-based decision-making through
inclusion of data quality. CEUR Workshop Proceed-
ings, 2491.
Deja, K. (2019). Using machine learning techniques for
data quality monitoring in cms and alice. Proceedings
of Science, 350.
Efron, B. (2020). Prediction, estimation, and attribu-
tion. Journal of the American Statistical Association,
Ehrlinger, L., Haunschmid, V., Palazzini, D., and Lettner,
C. (2019). A daql to monitor data quality in machine
learning applications. In Database and Expert Sys-
tems Applications, pages 227–237.
El Ioini, N., H
astbacka, D., Pahl, C., and Taibi, D. (2021).
Platforms for serverless at the edge: A review. In
Advances in Service-Oriented and Cloud Computing.
Fang, D., Liu, X., Romdhani, I., Jamshidi, P., and Pahl, C.
(2016). An agility-oriented and fuzziness-embedded
semantic model for collaborative cloud service search,
retrieval and recommendation. Future Gener. Comput.
Syst., 56:11–26.
Femminella, M. and Reali, G. (2019). Gossip-based mon-
itoring of virtualized resources in 5g networks. In
IEEE INFOCOM 2019 - IEEE Conference on Com-
puter Communications Workshops, INFOCOM Work-
shops 2019, pages 378–384. IEEE.
Gand, F., Fronza, I., Ioini, N. E., Barzegar, H. R., Azimi,
S., and Pahl, C. (2020). A fuzzy controller for self-
adaptive lightweight container orchestration. In In-
ternational Conference on Cloud Computing and Ser-
vices Science CLOSER.
Hong, C. and Varghese, B. (2019). Resource management
in fog/edge computing: A survey on architectures,
infrastructure, and algorithms. ACM Comput. Surv.,
Jamshidi, P., Sharifloo, A. M., Pahl, C., Metzger, A., and
Estrada, G. (2015). Self-learning cloud controllers:
Fuzzy q-learning for knowledge evolution. In 2015
International Conference on Cloud and Autonomic
Computing, pages 208–211. IEEE Computer Society.
Javed, A., Malhi, A., and Fr
amling, K. (2020). Edge
computing-based fault-tolerant framework: A case
study on vehicular networks. In Intl Wireless
Communications and Mobile Computing Conference,
Karkouch, A., Mousannif, H., Moatassime, H. A., and
Noel, T. (2016). Data quality in internet of things: A
state-of-the-art survey. Journal of Network and Com-
puter Applications, 73:57 – 81.
Lwakatare, L. E., Raj, A., Bosch, J., Olsson, H. H., and
Crnkovic, I. (2019). A taxonomy of software engi-
neering challenges for machine learning systems: An
empirical investigation. In Agile Processes in Soft-
ware Engineering and Extreme Programming - 20th
International Conference. Springer.
Mantri, D., Prasad, N. R., and Prasad, R. (2013). BHCDA:
bandwidth efficient heterogeneity aware cluster based
data aggregation for wireless sensor network. In Intl
Confe on Advances in Computing, Communications
and Informatics. IEEE.
Mendonca, N. C., Jamshidi, P., Garlan, D., and Pahl, C.
(2021). Developing self-adaptive microservice sys-
tems: Challenges and directions. IEEE Software,
Mohammedameen, I., Mkwawa, I., and Sun, L. (2019).
Follow-me prefetching for video streaming over mo-
bile edge computing networks. In 2019 IEEE Smart-
World, pages 1937–1942. IEEE.
Pahl, C. and Azimi, S. (2021). Constructing dependable
data-driven software with machine learning. IEEE
Softw., 38(6):88–97.
Pahl, C., Jamshidi, P., and Zimmermann, O. (2018). Archi-
tectural principles for cloud software. ACM Transac-
tions on Internet Technology (TOIT), 18(2):17.
Pomalo, M., Le, V. T., Ioini, N. E., Pahl, C., and Barzegar,
H. R. (2020). Service migration in multi-domain cel-
lular networks based on machine learning approaches.
In Intl Conf on Internet of Things: Systems, Manage-
ment and Security, IOTSMS 2020. IEEE.
Scolati, R., Fronza, I., Ioini, N. E., Samir, A., and Pahl,
C. (2019). A containerized big data streaming archi-
tecture for edge cloud computing on clustered single-
board devices. In Intl Conf on Cloud Computing and
Services Science.
Sheng, V. S., Provost, F. J., and Ipeirotis, P. G. (2008). Get
another label? improving data quality and data mining
using multiple, noisy labelers. In 14th SIGKDD In-
ternational Conference on Knowledge Discovery and
Data Mining, pages 614–622. ACM.
Sicari, S., Rizzardi, A., Miorandi, D., Cappiello, C., and
Coen-Porisini, A. (2016). A secure and quality-aware
prototypical architecture for the internet of things. In-
formation Systems, 58:43 – 55.
Sridhar, V., Subramanian, S., Arteaga, D., Sundararaman,
S., Roselli, D. S., and Talagala, N. (2018). Model
governance: Reducing the anarchy of production ml.
In USENIX Annual Technical Conference.
Taibi, D., Lenarduzzi, V., and Pahl, C. (2020). Microser-
vices Anti-patterns: A Taxonomy. Springer.
Tokunaga, K., Kawamura, K., and Takaya, N. (2016). High-
speed uploading architecture using distributed edge
servers on multi-rat heterogeneous networks. In IEEE
International Symposium on Local and Metropolitan
Area Networks, LANMAN 2016, pages 1–2. IEEE.
von Leon, D., Miori, L., Sanin, J., Ioini, N. E., Helmer, S.,
and Pahl, C. (2019). A lightweight container middle-
ware for edge cloud architectures. In Fog and Edge
Computing, pages 145–170. Wiley.
Wang, F., Zhang, M., Wang, X., Ma, X., and Liu, J. (2020).
Deep learning for edge computing applications: A
state-of-the-art survey. IEEE Access, 8:58322–58336.
Zhao, H., Yi, D., Zhang, M., Wang, Q., Xinyue, S., and Zhu,
H. (2019). Multipath transmission workload balanc-
ing optimization scheme based on mobile edge com-
puting in vehicular heterogeneous network. IEEE Ac-
cess, 7:116047–116055.
CLOSER 2022 - 12th International Conference on Cloud Computing and Services Science