ISO/IEC 25000 and AI Product Quality Measurement PerspectivesAndrea TrentaUNINFO UNI TCTechnical Committee Artificial IntelligenceTurinItaly412
In previous papers [10, 8, 9] we discussed ISO/IEC 25000 application when new quality measures are defined. The definition of product quality measures for ML is challenging, because of the huge number of algorithms and their implementations, that implies a huge number of measures, too. In continuity with papers above, and consistently with ISO standards, we show through examples of measures of ML accuracy and explainability, how to define practical ISO/IEC 25000 compliant product quality measures for AI. Moreover, the paper can be considered for the works in AI standardization area.
Policy makers, industries, and academia are
facing the problem of building trust in AI; in the
following we present a positive perspective, based
on the actual scientific and standardization
context, that can contribute to building trust in AI
through a quantitative approach. Then, the paper
focuses on open quality metric issues and
proposes a solution.
2. Standardization context in AI
Policy makers have addressed the issue of AI
trustworthiness mainly, but not only, to the
international standardization body ISO/IEC SC42
and to the European standardization body
CEN/CENELEC JTC21 that have in charge the
drafting of technical standards in support of
industry and of lawful rules. For the scope of this
paper, we consider, among the others, the
standards based on ISO 25000 series that define
or contribute to define product quality for an AI
product [
21
]. The assessment of product quality,
possibly together with the assessment of process
quality [
22
], will be performed in the near future
on voluntary or mandatory basis, in the former
case to promote trustworthiness in AI systems, in
the latter case to get compliance to rules [
23
]. In
the following, we focus on ML based AI systems
[
15
].
3. Quality models for AI
In the following we sum up the status of AI
product quality standardization and possible
future direction to move on.
Product quality is faced by SQuaRE project
that is described in ISO/IEC 25000 series and is
based on quality model and its measures.
In a very brief manner, there are 4 quality
models: (1) software, (2) data, (3) quality in use,
(4) service. Over each model is defined a set of
characteristics (for (1) reliability, defectivity, etc.
for (2) currentness, accuracy, consistency, etc. for
(3) usability, freedom from risk, etc. for (4)
responsiveness, IT service maintainability, etc.),
and in turn, over each of the characteristic are
defined basic measures, (e.g. the measure
‘number of duplicated items’ for characteristic
‘consistency’ in (2) or the measure ‘failure rate’
for characteristic ‘maturity’ in (1)).
[
17
] and in [
16
], where peculiar ML aspects, such
as bias and trustworthiness are addressed through
new specific characteristics like, but not only,
‘similarity’ for the former and ‘accuracy’ and
‘transparency’ for the latter.
Explainability and controllability, are also
dealt in
specific ISO/IEC
standards under
development.
4. Categorization
Firstly, it should be noted that generally is
easier to
design
data quality
measures that
software quality ones, because most of former are
influenced
simply
by
data
values
[
7
][
17
][
12
][
13
][
14
][
8
][
9
]
while the latter are
influenced by many context variables [
6
][
16
] and
so they are harder to measure; moreover, to get the
software
measures comparable, the
software
should be categorized. The need of software
categorization emerged since the early stage of
SQuaRE project and was addressed in [
24
], that,
among the purposes, includes the quality support
through the appropriate association and weight
between
type
of
software
and
quality
characteristics (e.g. for an home-banking software
it is important ‘accessibility’, for a defense or
medical software it is important ‘reliability’); this
association and its weight (see also [
26
]), allows
in turn the design of the relevant measures and
helps homogeneous product quality evaluation by
software category.
Applying the categorization approach to AI
requires
more
careful analysis than
non-AI
software, as further considerations are needed: for
example, the characteristic of ‘reliability’ should
lead the evaluation of software for x-ray image
processing,
but
the
characteristic
of
‘transparency’ or ‘explainability’ should lead the
evaluation of software for an x-ray automated
diagnosis instead.
2 for ‘algorithm’ it is intended the categorization of the code that
perform the task, e.g., for the classification task, the ‘algorithm’ can
be either a neural network, or a decision tree, or a support vector
machine, or other.
In this perspective, an overall quality score Qs
could be a sum of j-measurements Mij for each of
the Wi weighted i-characteristics selected for the
evaluation, and should be comparable with the
relevant benchmarks Bij:
function of
where:
The main issue for designing and applying
measures in a ML context seems the manifold of
implementations, more than 82000 according to
We can define the implementation I as a
1) I = I(method, algorithm(library, parameters),
training(dataset, process))
‘method’ is the high-level categorization, about
40 in [
18
] like decision tree, k-means clustering,
neural networks, and others
‘algorithm’ is the type of method2 (es. ResNet for
method=NN)
‘library’ contains the code to be invoked for
evaluation (see machine learning process in [
52
])
‘parameters’ are the configuration data of the
includes
and
Mij=Mij(method, algorithm(library,
parameters), training(dataset, process))
With those definitions, benchmark Bij is the best
value Mij for the time being (e.g., for a full year)
for the i-characteristic and the j-measure3 among
all the K implementations of Ik
In the function I, the argument ‘library’ specifies
the code or library that represents or simulates the
code to be measured. This approach follows the
global research community attempt to describe
performance of most of the papers through their
code, that is often available and public (see e.g.
GitHub that hosts Linux Foundation projects in
the category of Trusted and Responsible AI e.g.
[
28
], [
29
], [
30
], [
31
], [
32
], [
33
], [
34
], [
35
], [
36
],
[
37
], and others [
38
][
43
] that are relevant for
explainability metrics). It should be noted that
some of the biggest metric projects are led or
supported by big companies like Meta Research
[
27
], IBM [
42
], Microsoft Research [
43
]. In
addition to these resources, there are others like
Scikit-learn, and computing tools, like Matlab or
Wolfram, that have developed their own ML
libraries, mainly on the most consolidated
algorithms, for free or commercial use.
6. Use case 1: accuracy
In figure 1 below an example where the
icharacteristic is accuracy, and the j-measure is
based on multi-class classification metrics [
25
]
and calculated for the test dataset ImageNet for
various image classification algorithms (from
ZFNet to NFNet of the neural network method);
figure 1 shows the progress Bij of different
implementations since 2014 (points in grey are
non-top performing implementations for the
date).
See also [
25
] for comparison with ISO benchmark
definition.
3 in (4) the j-measure is supposed as scalar; if the j-measure is a vector
or a matrix, the expression (4) should be adapted.
As an example, the grey points of the figure 1
can be calculated repeating along the time the
measurement Mij, where the i-characteristic=
accuracy, and the j-measure Hamming loss (j=1)
where Ik is defined in (1).
To get measurements as homogeneous as
possible while grouping commonalities, it is
advisable that in (1) some variables are not
varying, for figure 1 they are:
‘method’
= Neural Network,
‘training dataset’ = Imagenet,
‘process’
‘library’
then 1) becomes:
= one-step training, and
= library_url,
1a)
I = I(algorithm(parameters))
With such assumptions, we can define a
kfamily measure, as a group of measures where any
measure belonging to a family differs from any
other of the same family only for the value of a
subset of variables of the relevant
implementation. An example of a k-family
measure is in table 1, where, for the measurement
function ‘Hamming loss’, the measures of the
same k-family share the same method, library, and
training; moreover, each family differs one from
the other for the algorithm and its parameters.
CCCC is the acronym of the characteristic
relevant for the measure
ML identifies the Machine Learning application
F is the number assigned to the measurement
function family
k is the number assigned to the measure of the G
family
For example, the family of measures of
accuracy trough F1 score of NNs trained with
dataset MNIST over a certain library can be
identified by ID = Accu-ML-5-k.
For the correct identification in case of more
detailed measurement function, the ID can be
further extended in the format [
4
][
10
]:
(6)
CCCC-ML-F-k-AA-v
In the following it is provided a second use
case of an ML measure relevant to the
subcharacteristic4 ‘explainability’.
As an example, but not the only one, in the
field of medicine, the questions to answer are: can
an AI automated x-ray diagnosis compete with a
professional diagnosis? Do the patient trust in the
AI outcome? Some researchers are discouraged
from answering such questions [
19
] and conclude
that “unless there are substantial advances in
explainable AI, we must treat these systems (AI
automated diagnosis system) as black boxes”.
Being out of the scope of this paper to define any
requirement towards clinical procedures or
protocols, but keeping in mind health
professionals’ concerns, in the following we
propose a possible design of an explainability ISO
25000 compliant measure.
Explainability [
39
][
40
] plays an important role
as it can help evaluation and risk assessment [
41
].
As a rough definition, explainability helps
humans to understand the work done by the ML
system and can be measured considering the more
salient features that influence the ML decision or
forecast; usually the metric for explainability is
based on the higher-scored features, measured
through numerical values or heatmap pixels. In
other words, it is attempted to explain decisions
(e.g. a classification outcome like ‘this is a dog’)
splitting the whole input data in smaller portions
(features) and permuting the input example or
altering it; those altered input data usually
produce an altered decision and allow to identify
which input alterations were most likely to change
the output decision. If the input data is an image,
for example by occluding one by one of the n
featured parts of the image, the explanation will
produce an heatmap that indicates the m<n image
parts that contributed the most to the decision.
Explainability measurements can be used in
conjunction with accuracy ones for assessing
purposes; for example, in [
47
] the measurement
function implements the F1 score for accuracy
measure, and the CAM heatmap for explainability
measure.
4 According to [
16
], explainability is not a tier-1 characteristic; in this
paper, ‘explainability’ is relevant to ‘transparency’ that in turn is a
sub characteristic of ‘usability’. For the sake of simplicity, in the
following it will be referred as a tier-1 characteristic.
The area5 in red in figure 2b is the area (2D
feature) of the chest that mainly driven the ML
system to the diagnosis of pneumonia from the
analysis of the x-ray image of figure 2a.
The same approach can be used for example to
explain why a credit request on behalf of an
enterprise is rejected by a ML decision support
system used by a bank. In this case it is useful to
understand which are the single items that more
contributed to the credit rejection decision; in this
example they are the enterprise financial health
indexes (features) like Book Value of Total Debt
(MVE_BVTD), Sales\Total assets (S_TA),
industry, and other features. The ML was trained
with the dataset of historical credit rating, that
contained the decisions made on past lending
requests based on financial indexes of requesting
enterprises. A typical way to measure such
contributions is calculation of Shapley values
[
48
]. Following the previous use case and
definition (1), we can define an example of the
5 The explanation measurement in figure 2b is an heatmap matrix and
not a scalar numeric value. At the same manner, the measurement
function X in table 2 is a vector (1xF).
measurement Mij, where the i-characteristic=
explainability, and the j-measure is Shapley
values (j=1) and method and algorithm in (1) are
not referred to the original ML model but to the
simulated one, the so-called ‘explanation model’:
Again, the expression (1a) holds, as we have
chosen in table 2 to group by k the algorithms and
parameters as there are more than one ‘Shapley
values’ possible ways of calculation.
The influence of training data G and O on the
regression parameters of the machine, suggests
that even the description of the measure
Expl-ML1-k shall be intended as a general measure that is
not suitable for practical purposes; therefore, the
measures in table 2 should be further detailed
through:
i. a new ID in the format (6),
ii. the specification of the domain D allowed
for query points (as the explanation
model works fine only for local query
points).
Under the same code of the model and the
same training dataset and training process, the
specific measure can be applied in practice for all
the query points in the local defined domain. Such
a further specification is needed, for example, to
check if the ML is ‘explainable’ for an enterprise
that belongs to an industry that was not (outlier)
in the training data (as the enterprise financial
indexes are the query point).
8. Proposal
To establish a framework of meaningful and
comparable measures for AI applications requires
a new approach: the manifold issue compels us to
define as many measures and specific benchmarks
as the thousands of tasks, algorithms, dataset
combinations are. This is why general measures
would be hard to practice and would be not
comparable.
Then, consistently with the structure of
libraries available from the AI research
community, we derived the definition (2)
Mij=Mij(I), that is applied both in the example in
Table 1 for accuracy and in Table 2 for
explainability. Both examples represent the
proposal: it consists in a detailed product quality
measure design and documentation that includes
algorithm, training dataset, library code and
parameters; moreover, it was considered the
chance to group by family homogeneous
measures. Last but not the least, the proposal is
conceived to be compliant to ISO/IEC 25000.
9. Conclusion
The spread of AI applications in fields like
finance, healthcare, transportation, urges to build
trustworthiness in users; policy makers are facing
this issue and so developing several norms, e.g.
[
49, 50, 51
].
The contribution of ISO 25000 models and
measures to the construction of a trusted AI
environment is already recognized [
16
], but it will
be maximally effective only if the measurements
will be appropriately assessed, and benchmarks
will be available. A way to do this, is through the
approach proposed that in turn is strictly based on
the actual categorization and organization of AI
libraries developed by the research community
[
27-38, 42, 43
] and inspired by successful similar
experiences in creating listed measures [
46
].
The present proposal, as well as the measure
naming and process described in [
10
], could be
considered by the ISO SC42 and SC7 relevant
working groups for the standardization work in
progress.
10.
[1]
InternationalOrganization
for Standardization, ISO/IEC 25000:
2014 Systems and Software engineering - Systems and software Quality Requirements and Evaluation (SQuaRE) - Guide to SQuaRE
. URL: https://standards.iso.org/ittf/PubliclyAvailab leStandards/index.html
[2]
InternationalOrganization
for Standardization, ISO/IEC 25010:
2011 Systems and Software engineering - Systems and software Quality Requirements and Evaluation (SQuaRE) - System and software quality models
. URL: https://www.iso.org/standard/35733.html
[3]
InternationalOrganization
for Standardization,
ISO/IEC 25012:2008 Systems and Software engineering - Systems and software Quality Requirements
and
Evaluation (SQuaRE
)
- Data quality model
. URL: https://www.iso.org/standard/35736.html
[4]
InternationalOrganization
for Standardization, ISO/IEC 25020:
2019
,
Systems and Software engineering - Systems and software Quality Requirements and Evaluation (SQuaRE) - Quality measurement framework
. URL: https://www.iso.org/standard/72117.html
[5]
InternationalOrganization
for Standardization, ISO/IEC 25022:
2016
,
Systems and Software engineering - Systems and software Quality Requirements and Evaluation (SQuaRE) - Measurement of quality in use
. URL: https://www.iso.org/standard/35746.html
[6]
InternationalOrganization
for Standardization, ISO/IEC 25023:
2016
,
Systems and Software engineering - Systems and software Quality Requirements and Evaluation (SQuaRE) - Measurement of system and software product quality
. URL: https://www.iso.org/standard/35747.html
[7]
InternationalOrganization
for Standardization, ISO/IEC 25024:
2015
,
Systems and Software engineering - Systems and software Quality Requirements and Evaluation (SQuaRE) - Measurement of data quality
. URL: https://www.iso.org/standard/35749.html
[8]
A.Trenta
,
Data bias measurement: a geometrical approach through frames
,
Proceedings of IWESQ@APSEC2021
. URL: http://ceur-ws.
org/
Vol-
3114
/
[9]
A.Trenta: ISO/IEC 25000 quality measures for A.I.: a geometrical approach
,
Proceedings of IWESQ@APSEC2020
. URL: http://ceur-ws.
org/
Vol-
2800
/
[10]
D.Natale
,
A.Trenta
,
Examples of practical use of ISO/IEC 25000, Proceedings of IWESQ@APSEC2019
. URL: http://ceurws.org/Vol-
2545
/
[11] International Organization for Standardization,
ISO/IEC TR 24027 Information technology - Artificial Intelligence (AI
)
- Bias in AI systems and AI-aided decision making
. URL: https://www.iso.org/standard/35749.html
[12]
M.Mecati
,
F. E.Cannavò
,
A.Vetrò
,
M.Torchiano
,
Identifying Risks in Datasets for Automated Decision-Making
, in: International Conference on Electronic Government,
2020
, pp.
332
-
344
, Springer,Cham. URL: https://link.springer.com/chapter/10.1007%
2F978
-
3
-
030
-57599-1_
25
[13]
E.Beretta
,
A.Vetrò
,
B.Lepri
,
J.C. De Martin
,
Detecting discriminatory risk through data annotation based on Bayesian inferences
,
in: Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency
,
2021
, pp.
794
-
804
. URL: https://doi.org/10.1145/3442188.3445940
[14]
E.Beretta
,
A.Vetrò
,
B.Lepri
,
J.C. De Martin
,
Ethical and socially-aware data labels
,
in Annual International Symposium on Information Management and Big Data
,
2018
, pp.
320
-
327
, Springer, Cham. URL: https://link.springer.com/chapter/10.1007%
2F978
-
3
-
030
-11680-4_
30
[15] International Organization for Standardization,
ISO/IEC 22989:2022 Information technology - Artificial intelligence -Artificial intelligence concepts
and
terminology
. URL: https://www.iso.org/standard/74296.html
[16]
International Organization for Standardization, ISO/IEC DIS 25059 Software engineering - Systems and software Quality Requirements and Evaluation (SQuaRE) - Quality Model for AI-based systems
. URL: https://www.iso.org/standard/80655.html
[17]
International Organization for Standardization, ISO/IEC CD 5259-2 (under development) Artificial intelligence - Data quality for analytics and ML - Part 2: Data quality measures
. URL: https://www.iso.org/standard/81860.html
[18] International Organization for Standardization, ISO/IEC 23053:
2022Framework for Artificial Intelligence (AI) Systems Using Machine Learning (ML)
. URL: https://www.iso.org/standard/74438.html
[19]
M.Ghassemi
,
L.Oakden-Rayner
,
A. L.Beam
,
The false hope of current approaches to explainable artificial intelligence in health care Lancet Digit Health
,
2021
,
3
:
e745
-
50
. URL: https://www.thelancet.com/pdfs/journals/lan dig/PIIS2589-
7500
(
21
)
00208
-
9
.pdf
[20]
M.Chetoui
,
M.A.Akhloufi
,
B.Yousefi
,
E.M.Bouattane
,
ExplainableCOVID-19 Detection on Chest X-rays Using an End-toEnd Deep Convolutional Neural Network Architecture
,
Big Data and Cognitive Computing
,
2021
,
5
,
73
. URL: https://doi.org/10.3390/bdcc5040073
[21]
D.Natale
,
Extensions of ISO/IEC 25000 quality models to the context of Artificial Intelligence
,
Proceedings of IWESQ@APSEC 2022
. To appear.
[22]
International Organization for Standardization, ISO/IEC 42001 (draft) Information technology - Artificial intelligence - Management system
. URL: https://www.iso.org/standard/81230.html
[23]
EuropeanCommission
, COM/
2021
/206 '
Proposal for a regulation of the european parliament and of the council laying down harmonised rules on artificial intelligence (artificial intelligence act) and amending certain union legislative acts
',
2021
. URL: https://eur-lex.europa.eu/legalcontent/EN/TXT/?uri=
CELEX:52021PC020 6
[24] International Organization for Standardization, ISO/IEC TR 12182
:2015 Systems and software engineering - Framework for categorization of IT systems and software, and guide for applying it
. URL: https://www.iso.org/standard/63611.html
[25] International Organization for Standardization, ISO/IEC TS 4213:2022
InformationTechnology - Artificial
Intelligence -
Assessment of machine learning classification performance
. URL: https://www.iso.org/standard/79799.html
[26]
International Organization for Standardization, ISO/IEC CD 25040 (under development) Systems and software engineering - Systems and software Quality Requirements and Evaluation (SQuaRE) - Quality evaluation framework
. URL: https://www.iso.org/standard/83467.html
[27]
MetaResearch
,
Papers with Code resource
. URL: https://paperswithcode.com/sota
[28]
Trusted-AI
,
Contrastive explanation methods resource
. URL: https://github.com/IBM/AIX360/tree/master /aix360/algorithms/contrastive/CEM.py for A.
Dhurandhar
et al.,
Explanations based on the Missing: Towards Contrastive Explanations with Pertinent Negatives, Advances in Neural Information Processing Systems (NeurIPS)
,
2018
. https://arxiv.org/abs/
1802
.07623
[29]
S.Lundberg
,
SHAP resource
. URL: https://github.com/slundberg/shap for Lundberg, et al.
A Unified Approach to Interpreting Model Predictions
,
Advances in Neural Information Processing Systems (NeurIPS)
,
2017
. URL: http://papers.nips.cc/paper/7062-a
-unifiedapproach-to-interpretingmodel-predictions
[30]
Trusted-AI
,
Contrastive Explanations Method with Monotonic Attribute Functions resource
. URL: https://github.com/TrustedAI/AIX360/blob/master/aix360/algorithms/c ontrastive/CEM_MAF.py for R. Luss et al.,
Leveraging Latent Features for Local Explanations
,
2019
, URL: https://arxiv.org/abs/
1905
.12698
[31]
Trusted-AI
,
Explainability resource
. URL: https://github.com/TrustedAI/AIX360/blob/master/aix360/algorithms/t ed/TED_Cartesian.py for M.
Hind
et al.,
TeachingAI
to
Explain its Decisions
,
Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society
,
2019
, URL: https://doi.org/10.1145/3306618.3314273
[32]
Trusted-AI
,
Faithfulness resource
. URL: https://github.com/TrustedAI/AIX360/blob/master/aix360/metrics/loca l_metrics.py, for D.
Alvarez-Melis
,
T.Jaakkola
,
Towards Robust Interpretability with Self-Explaining Neural Networks
,
Advances in Neural Information Processing Systems31
(
NeurIPS
),
2018
, URL: https://papers.nips.cc/paper/8003-towardsrobust
-interpretabilitywith-self-explainingneural-networks
[33]
Trusted-AI
,
Monotonicity resource
. URL: https://github.com/TrustedAI/AIX360/blob/master/aix360/metrics/loca l_metrics.py for R. Luss et al.,
Leveraging Latent Features for Local Explanations
,
2019
, URL: https://arxiv.org/abs/
1905
.12698
[34]
Trusted-AI
,
ProfWeight resource
. URL: https://github.com/TrustedAI/AIX360/blob/master/aix360/algorithms/ profwt/profwt.py for A.
Dhurandhar
et al.,
Improving Simple Models with Confidence Profiles, Advances in Neural Information Processing Systems31
(
NeurIPS
),
2018
, URL: https://papers.nips.cc/paper/8231- improving
-simple-models-withconfidenceprofiles
[35]
Trusted-AI
,
ProtoDash resource
. URL: https://github.com/TrustedAI/AIX360/blob/master/aix360/algorithms/ protodash/PDASH.py for S. Gurumoorthy et al.,
Efficient Data Representation by Selecting Prototypes with Importance Weights
,
International Conference on Data Mining (ICDM)
,
2019
, URL: https://arxiv.org/abs/1707.01212
[36]
Trusted-AI
,
Boolean Decision Rules via Column Generation resource
. URL: https://github.com/TrustedAI/AIX360/blob/master/aix360/algorithms/r bm/BRCG.py
(Light Edition) for S. Dash et al
.,
Boolean Decision Rules via Column Generation
,
Advances in Neural Information Processing Systems31
(
NeurIPS
),
2018
, URL: https://papers.nips.cc/paper/7716- boolean
-decision-rules-via-columngeneration
[37]
Trusted-AI
,
Generalized Linear Rule Models resource
. URL: https://github.com/TrustedAI/AIX360/blob/master/aix360/algorithms/r bm/GLRM.py for D.
Wei
et al.,
Generalized Linear Rule Models, Proceedings of the 36th International Conference on Machine Learning
, PMLR
97
:
6687
-
6696
,
2019
, URL: http://proceedings.mlr.press/v97/wei19a.ht ml
[38]
H. X.Vinh
,
Quantitative Input Influence resource
. URL: https://github.com/hovinh/QII, for A.
Datta S.Sen
,
Y.Zick
, Algorithmic Transparency via Quantitative Input Influence:
Theory and Experiments with Learning Systems
,
IEEE Symposium on Security and Privacy (SP)
,
2016
. URL: https://ieeexplore.ieee.
org/document/754652 5
[39]
International Organization for Standardization, ISO/IEC AWI TS 6254 (under development) Information technology - Artificial intelligence - Objectives and approaches for explainability of ML models and AI systems
. URL: https://www.iso.org/standard/82148.html
[40] International Organization for Standardization,
ISO/IEC TR 24028:2020 Information technology -Artificial intelligence (AI
)
- Overview of trustworthiness in artificial intelligence
. URL: https://www.iso.org/standard/77608.html
[41]
International Organization for Standardization, ISO/IEC CD TR 5469 (under development) Artificial intelligence - Functional safety and AI systems
. URL: https://www.iso.org/standard/81283.html
[42]
AIExplainability
360 - IBM Research Trusted AI URL: https://aix360.mybluemix.net/
[43]
MicrosoftResearch,InterpretML toolkit resource
. URL: https://interpret.ml
[44]
M.Sameki
,
S.Bird
,
K.Walker
, et al.
InterpretML: A toolkit for understanding machine learning models
,
2020
. URL: https://www.microsoft.com/enus/research/uploads/prod/2020/05/
InterpretML
-Whitepaper.pdf
[45]
Scikit-learn resource
. URL: https://scikitlearn.org/stable/index.html#
[46] International Organization for Standardization,
ISO/IEC 5055:2021 Information technology - Software measurement - Software quality measurement - Automated source code quality measures
. URL: https://www.iso.org/standard/80623.html
[47]
P.Rajpurkar
,
J.Irvin
,
K.Zhu
, et al.
CheXNet: radiologist-level pneumonia detection on chest X-rays with deep learning, 2017
. URL: http://arxiv.org/abs/1711.05225
[48]
Mathworks resource
. URL: https://it.mathworks.com/help/stats/shapleyvalues
-for-machine-learning-model
.html
[49]
FederalReserveBoard
, Equal Credit Opportunity Act, URL: https://www.federalreserve.gov/boarddocs/s upmanual/cch/fair_lend_reg_b.pdf
[50]
ConsumerFinancial Protection Bureau
, Fair Credit Reporting Act, URL: https://www.consumerfinance.gov/rulespolicy/regulations/1022/
[51]
EuropeanCommission
,
Proposal for a 'Regulation of the european parliament and of the council laying down harmonised rules on artificial intelligence (artificial intelligence act) and amending certain union legislative acts' COM/2021
/206 final URL: https://eur-lex.europa.eu/legalcontent/EN/TXT/?uri=
CELEX:52021PC020 6
[52] International Organization for Standardization, ISO/IEC 23053:
2022Framework for Artificial Intelligence (AI) Systems Using Machine Learning (ML)
. URL: https://www.iso.org/standard/74438.html