ML System Engineering Supported by a Body of Knowledge

Juliette Mattioli

, Dominique Tachet

, Fabien Tschirhart

, Henri Sohier

, Loic Cantat

and

Boris Robert

2,3

Thales, France

IRT SystemmX, France

IRT Saint Exup

ery, France

Keywords:

Body-of-Knowledge, Knowledge Graph, Knowledge Extraction, Knowledge Fusion, ML Engineering.

Abstract:

A body of knowledge (BoK) can be deﬁned as the comprehensive set of concepts, terminology, standards, and

activities that facilitate the dissemination of knowledge about a speciﬁc ﬁeld, providing guidance for practice

or work. This paper presents a methodology for the construction of a body of knowledge (BoK) based on

knowledge-based artiﬁcial intelligence. The process begins with the identiﬁcation of relevant documents and

data, which are then used to capture concepts, standards, best practices, and state-of-the-art. These knowledge

items are then fused into a knowledge graph, and ﬁnally, query capacities are provided. The overall process

of knowledge collection, storage, and retrieval is implemented with the objective of supporting a trustworthy

machine learning (ML) end-to-end engineering methodology, through the ML Engineering BoK.

1 RATIONALE

Systems and products are developed in competitive,

volatile, uncertain, complex and ambiguous contexts,

inﬂuenced by external factors such as regulations

and societal expectations. Artiﬁcial Intelligence (AI)

technologies, in particular Machine Learning (ML)

approaches, improve product quality and production

efﬁciency (Li et al., 2017). Reliability is crucial for

critical systems to remain reliable throughout their

lifecycle and to evolve cost-effectively. However,

the rapid adoption of AI technologies is leading to

a specialization of engineers and a dispersion of the

required skills. Current demographics are making

highly skilled and experienced engineers scarce, lead-

ing to a lack of project support and mentorship. In ad-

dition, the complexity of AI-based solutions and past

trends in each component require the management of

AI engineering knowledge and general engineering

practices. This highlights the need for effective man-

agement of AI engineering knowledge and practices.

Knowledge management and knowledge engineering

(KE) are often used interchangeably, with ”manage”

referring to executive leadership and ”engineer” to

planning, construction, or design activities. The main

difference is that the knowledge manager sets the

process direction, while the knowledge engineer de-

velops the means to achieve it. The Conﬁance.ai

program’s end-to-end methodology serves as a foun-

dational framework for knowledge management in

trustworthy AI engineering (Awadid et al., 2024). It

addresses non-functional requirements for successful

implementation of ML-based components in critical

systems (Adedjouma et al., 2022). The methodol-

ogy covers various process levels and aligns with in-

dustrial best practices. It is essential to deﬁne the

scope and position of the methodology in relation to

other engineering disciplines. KE is a sub-ﬁeld of AI

that focuses on understanding, designing, and imple-

menting methods for representing information effec-

tively (Shapiro, 2006). It facilitates the management

of all types of knowledge based on labelled graphs

and provides guidance for resolving ML engineering-

related issues. In order to address these issues, Conﬁ-

ance.ai program put forth the argument that there is a

need to consolidate the AI engineering ﬁeld’s largely

fragmented body of knowledge (Mattioli et al., 2024),

which encompasses data engineering, algorithm engi-

neering, software and system engineering, safety and

cyber-security, similar to the SWEBOK deﬁnition of

software engineering (Robert et al., 2002).

www.conﬁance.ai/en

Mattioli, J., Tachet, D., Tschirhart, F., Sohier, H., Cantat, L. and Robert, B.

ML System Engineering Supported by a Body of Knowledge.

DOI: 10.5220/0013048500003838

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 16th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2024) - Volume 3: KMIS, pages 331-338

ISBN: 978-989-758-716-0; ISSN: 2184-3228

331

Figure 1: AI/ML deployment induces some (engineering) challenges.

2 KNOWLEDGE ENGINEERING

From this perspective, Knowledge Engineering (KE)

represents a sub-ﬁeld of AI that is concerned with the

understanding, design and implementation of meth-

ods for representing information in a manner that

enables computers to utilize it effectively. In other

words, the objective of KE is the understanding and

subsequent representation of human knowledge in the

form of data structures, semantic models (conceptual

graph of the data in relation to the real world) and

heuristics (or rules). The induced process comprises

three principal elements: a) knowledge acquisition,

representation, and validation; b) inference; and c)

explanation and justiﬁcation. It is evident that ap-

proaches to knowledge representation, such as on-

tologies, taxonomies, thesauri, and vocabularies, can

be employed to support a diverse array of activities.

Techniques that could be used for such purpose in-

clude:

• Using a taxonomy that is well-known and ac-

cepted by the organization in work-products;

• Using patterns to help a computer “understand”

the content of work-products;

• Encoding the breakdown structures that represent

well-established knowledge within the organiza-

tion;

• Using inference rules to reason about the quality

of the content of different work-products.

An organization must manage knowledge about its

systems and engineering practices, known as a Body

of Knowledge (BoK). The BoK provides a compre-

hensive description of ML system engineering con-

tents and practices, establishing a foundation for cur-

riculum development and creating a coherent curricu-

lum for qualiﬁcation towards certiﬁcation.

3 ML ENGINEERING BoK

A body of knowledge (BoK) is”structured knowl-

edge that is employed by members of a discipline to

inform their practice or work” (

Oren, 2005). It is used

to deﬁne concepts and activities, such as accuracy

in data engineering, machine learning models and

system-level applications. The core components of a

BoK include concepts, knowledge, skills, standards,

terminology, guidelines, practices and activities. It

serves as the ”ground truth” for ML engineering ac-

tivities, covering various domains such as data engi-

neering, algorithm engineering, software engineering,

systems engineering, cyber-security, safety, and cog-

nitive engineering.

3.1 The BoK Design

The BoK design is an iterative process involv-

ing knowledge acquisition, fusion, storage, and re-

trieval. Knowledge is acquired from structured, semi-

structured, and unstructured data, with extraction fo-

cusing on entities, attributes, and relations. Knowl-

edge fusion requires ongoing ontology construction

and quality evaluation. Currently, knowledge is typ-

ically stored in KG databases. Conﬁance.ai BoK

(ﬁg 2) is a comprehensive guide for engineers on

the lifecycle of AI-based critical systems. It offers

guidance and support throughout the development,

maintenance, and evolution of these systems. The

guide deﬁnes trustworthy ML engineering concepts

and provides an outline of essential knowledge, skills,

KMIS 2024 - 16th International Conference on Knowledge Management and Information Systems

332

Figure 2: Trustworthy ML engineering Body-of-Knowledge - https://bok.conﬁance.ai/.

and practices, covering all fundamental competencies

required by professionals in the ﬁeld. The initial step

of the ML Engineering BoK design is to identify the

domain of application and compile a list of relevant

knowledge sources. Secondly, a conceptual model

will be devised with the objective of gathering to-

gether the entities of interest, their inter-relationships

and the categories. A valuable resource for concep-

tual modeling is Capela©, which is a model-based en-

gineering solution that has proven effective in numer-

ous industrial contexts. Thirdly, the logical and phys-

ical models will provide a logical representation and

assertions for the entities and relationships that have

been collected. Fourthly, the technical development

and implementation must take into account the cod-

ing language to be employed (for example, RDF and

OWL), as well as the serialization formats (such as

RDF/XML, Turtle and JSON-LD). The ﬁnal stage is

the deployment of the BoK as a service, thereby facil-

itating reuse and enabling the engineering community

to provide feedback. In essence, this process entails

the transformation of knowledge held by engineering

experts and end-users into a machine-readable format.

Designing a ML Engineering BoK induces some gen-

eral issues:

• Raw Data Acquisition. How to select the rele-

vant data and information to be fed into the BoK?

• Knowledge Extraction, Representation and

Validation. How do we represent human knowl-

edge as it currently exists in state of the art reports,

scientiﬁc articles, standards and norms, Engineer-

ing best practices, and the minds of the experts in

terms of data structures that can be processed by

a computer? How to determine the best represen-

tation for any given engineering problem?

• Knowledge Integration and Fusion. How do we

use these knowledge item to generate useful infor-

mation in the context of a ML engineering?

• KG Design. How to manipulate the knowledge to

provide explanations to the engineer/user?

• Knowledge Query. How do we use these abstract

knowledge structures to generate useful informa-

tion in the context of a speciﬁc case?

3.2 Step 1: Raw Data Acquisition

BoK developers create knowledge bases from scratch,

dealing with diversity and heterogeneity of knowl-

edge representation formalisms and mismatch of dif-

ferent knowledge items. Knowledge engineers focus

on modeling structural use cases and expert knowl-

edge concepts. The ﬁrst step is to deﬁne a taxonomy

of the ML engineering domain aligned with ISO/IEC

DIS 5338 standard for AI systems and safety and reli-

ability standards. This taxonomy is the classiﬁcation

of concepts induced by Trustworthy ML Engineering

activities. The operational deﬁnition of trustworthy

AI includes a taxonomy and keywords that deﬁne core

domains such as AI Engineering, Data Engineering,

and Safety Engineering, covering the entire life cycle

of AI-based critical systems. This framework aids in

harmonizing design and support activities, including

monitoring and maintenance, and serves as a founda-

tion for the Conﬁance.ai methodology, which outlines

requirements and recommendations. The initial stage

is devoted to the identiﬁcation of data sources, as this

ML System Engineering Supported by a Body of Knowledge

333

has a signiﬁcant impact on the entire knowledge graph

(KG) development process, as well as on the selection

of knowledge extraction techniques. Deﬁnitions have

been collated from various sources, including Euro-

pean and worldwide standardization bodies (ISO/IEC

5338, Aerospace Standard 6983, IEEE 7000...), Na-

tional and European projects (Conﬁance.ai, DEEL

project

, JRC Flagship on AI

), scientiﬁc publica-

tions and working groups (e.g. the HLEG or the AI

Safety Landscape initiative

), as well as other relevant

sources. The working group responsible for the state

of the art sourced deﬁnitions from external literature

in most cases. Sometimes the existing literature did

not match the scope of ML Engineering. We created

a new deﬁnition. This phase was also based on data

about making AI reliable. This includes making AI

reliable through design, data engineering for trusted

AI, IVVQ Strategy (Integration, Veriﬁcation, Valida-

tion and Qualiﬁcation), and targeted embedded AI.

3.3 Step 2: Knowledge Extraction,

Representation and Validation

The extraction of knowledge from semi-structured

sources is easier than from unstructured sources,

which hold more information. The second phase fo-

cuses on extracting knowledge from unstructured data

to create and enhance knowledge graphs, identify-

ing entities and relationships. This process involves

natural language processing (NLP) and knowledge

representation technologies to automatically extract

structured information from various data types, fa-

cilitating the effective use of external data. An en-

tity represents the most fundamental unit of a knowl-

edge graph. It represents a concept. Furthermore,

the quality of knowledge graph construction is contin-

gent upon the accuracy and integrity of its extraction.

Subsequently, relationship extraction entails the iden-

tiﬁcation of associations between entities, thereby es-

tablishing semantic relations and forming a knowl-

edge network. These types of graphs embed a struc-

tured representation of facts, consisting of entities,

relationships, and semantic descriptions, which are

modeled with an RDF (Resource Description Frame-

work) structure. An RDF model is a ﬂexible data rep-

resentation model comprising three-element tuples,

with no ﬁxed schema requirements. It is a graph-

based model for the description of entities and their

relationships on the Web. Many researchers prefer

https://www.deel.ai/

https://joint-research-centre.ec.europa.eu/jrc-mission-

statement-work-programme/facts4eufuture/artiﬁcial-

intelligence-european-perspective/future-ai en

https://www.aisafetyw.org/ai-safety-landscape

to conceptualize RDF as a set of triples, although

it is commonly described as a directed and labelled

graph, each consisting of a subject, predicate and ob-

ject in the form of < sub ject, predicate, ob ject >. In

this context, the predicate represents the relationship

between the subject and the object. For example:

< Data Engineering, is an activity, MLOps >. The

triples are stored in a triple store and can be queried

using the SPARQL query language. In comparison

to both inverted indices and plain text ﬁles, triple

stores and the SPARQL query language facilitate the

formulation of sophisticated queries, enabling users

to satisfy complex information needs. Although a

model is required for representing data in triples (sim-

ilar to relational databases), RDF enables the expres-

sion of rich semantics and supports knowledge infer-

ence (Hertz et al., 2019). Like any model, such a

BoK is only an approximation of reality. New ob-

servations based on ML engineering use-cases can

guide the further acquisition of knowledge. There-

fore, an evaluation of the represented knowledge with

respect to reality is indispensable for the creation of

an adequate model. These limitations relate to the

so-called symbol grounding problem (Harnad, 1990),

and concern the extent to which representational ele-

ments are hand-crafted rather than learned from data.

The most common methods employed include pat-

tern matching, machine learning, and semantic rule

extraction. Furthermore, generative models such as

large language models (LLMs) can play a pivotal

role in the construction of knowledge graphs by ex-

tracting entities, relationships, and attributes from un-

structured text data (Meyer et al., 2023). They can

be pre-trained through the structurally consistent lin-

earization of text, which facilitates the transition from

traditional understanding to structured understanding

and increases knowledge sharing (Wang et al., 2022).

In contexts with Named Entity Recognition (NER),

as demonstrated by (Strakov

a et al., 2019), the pro-

posed generative method implicitly models the struc-

ture between named entities. This approach effec-

tively avoids the complexity inherent to multi-label

mapping. Similarly, extracting overlapping triples in

relation extraction is also challenging to address for

traditional discriminating models (Zeng et al., 2018),

introducing a new perspective for addressing this is-

sue through a general generative framework. More-

over, several features must be taken into account when

developing a BoK:

• Redundancy: Are there identical or equivalent

knowledge items that is a special case of another

(subsumed)?

• Consistency: Are there ambiguous or conﬂict-

ing knowledge, is there indeterminacy in its ap-

KMIS 2024 - 16th International Conference on Knowledge Management and Information Systems

334

plication? Is it intended? Are several outcomes

possible, for example, depending on the strategy

(the order in which the knowledge models are or-

dered)?

• Minimality: Can the knowledge set be reduced

and simpliﬁed? Is the reduced form logically

equivalent to the ﬁrst one?

• Completeness: Are all possible entries covered by

the knowledge of the set?

Thus, a good BoK must have properties such as:

• Representational Accuracy: It should represent all

kinds of required knowledge.

• Inferential Adequacy: It should be able to ma-

nipulate the representational structures to pro-

duce new knowledge corresponding to the exist-

ing structure.

• Inferential Efﬁciency: The ability to direct the in-

ferential knowledge mechanism into the most pro-

ductive directions by storing appropriate guides.

• Acquisitional Efﬁciency: The ability to complete

with new knowledge easily using automatic meth-

ods.

Peer reviews with various stakeholders (data scien-

tists, software and system engineers, safety and cyber-

security engineers...) were carried out to assess the

appropriateness and quality of the acquired knowl-

edge in relation to the ML engineering end-to-end

methodology (Adedjouma et al., 2022).

3.4 Step 3: Knowledge Integration and

Fusion

For (Sowa, 2000), “Knowledge Representation is the

application of logic and ontology to the task of con-

structing software models for some domain”. There-

fore, the way a knowledge representation is con-

ceived reﬂects a particular insight or understanding

of how people reason. The selection of any of the

currently available representation technologies (such

as logic, knowledge bases, ontology, semantic net-

works...) commits one to fundamental views on the

nature of intelligent reasoning and consequently very

different goals and deﬁnitions of success. As we ma-

nipulate concepts with words, all ontologies use hu-

man language to ”represent” the world. Thus, ontol-

ogy is expressed as a formal representation of knowl-

edge by a set of concepts within a domain and the

relationships between these concepts. Nevertheless,

the ”ﬁdelity” of the representation depends on what

the knowledge-based system captures from the real

thing and what it omits. If such system has an im-

perfect model of its universe, knowledge exchange

or sharing may increase or compound errors during

the ML Engineering process. As such, a fundamen-

tal step is to establish effective knowledge represen-

tation (symbolic representation) that can be used for

query. The sheer complexity, variety and volume of

data available today presents a signiﬁcant challenge

to achieving efﬁcient and accurate knowledge graph

fusion. The process of integrating disparate sources

of knowledge, also known as knowledge fusion, en-

tails the elimination of redundancies, inconsistencies,

and ambiguities from the integrated corpus. The ﬁeld

of engineering is one in which knowledge is typically

subject to updates. In the majority of cases, users

will have the capacity to supplement existing exter-

nal knowledge graphs with external knowledge. Thus,

the objective of knowledge fusion is to merge seman-

tically equivalent elements, for example, the concepts

of ”accuracy” and ”machine learning accuracy”, with

the intention of integrating novel forms of knowledge

within existing conceptual frameworks or factual as-

sertions. The sub-tasks of knowledge fusion include

the alignment of attributes, the matching of entities

with small-scale incoming triples, and the alignment

of entities with a complete knowledge graph. This

stage is beneﬁcial for both the generation and com-

pletion of knowledge graphs. We employ the knowl-

edge graph representation for knowledge fusion, as

proposed by (Laudy et al., 2007), which is based on

the conceptual graph model. This representation is

used to store and combine knowledge. The approach

is to examine observations with domain knowledge

and graph operators. This removes any bias from

translating data from one format to another with dif-

ferent models. We suggest using it for a high-level in-

formation fusion approach based on the Maximal Join

operator, which is an aggregation operator on concep-

tual graphs (Laudy, 2011).

3.5 Step 4: Knowledge Graph

At worst, the effort involved in specifying the rele-

vant knowledge forces us to think more deeply about

the relevant ways of characterizing the ML engineer-

ing models that we as researchers implicitly construct

anyway. The use of Knowledge Graphs (KG) as

a means of representing knowledge is becoming in-

creasingly prevalent. Their versatility in terms of rep-

resentation allows for the integration of diverse data

sources, both within and across engineering bound-

aries. Therefore, our primary strategy to support

this step in practice was the creation of a knowledge

graph, which collects information about each ML de-

velopment activity, the artifacts and processes used

in the entire ML-based system lifecycle, the end-to-

end methodology, and the motivation behind the key

ML System Engineering Supported by a Body of Knowledge

335

design decisions. Several replications have been car-

ried out in this way, contributing to a growing body

of knowledge about ML engineering techniques. By

employing the graph architecture, KGs are capable of

modeling a range of relationship types (edges) and

entities (nodes) (Chen et al., 2020). KGs comprise

an additional embedded layer, designated a reasoner

(or inference engine), which enables them to extract

implicit information from existing explicit concepts,

in contrast to plain graph or non-relational databases.

The most well-known examples of knowledge graphs

(KGs) – DBpedia, Freebase, Wikidata, YAGO, and so

forth – encompass a diverse array of domains and are

either derived from Wikipedia or created by volunteer

communities (Heist et al., 2020). The Google Knowl-

edge Graph is one of the largest and most comprehen-

sive KGs in existence, aiming to model and link all

structured information found on the internet, includ-

ing persons, organizations, skills, events, products,

and more. This is one of the reasons why the Google

search engine is so effective. A graph-based knowl-

edge representation and reasoning formalism derived

from conceptual graphs has been formalized as ﬁnite

bipartite graphs, as outlined in (Mugnier and Chein,

1992). In this formalism, the set of nodes is divided

into concept and conceptual relation nodes. In such

a graph, concept nodes represent classes of individu-

als, and conceptual relation nodes illustrate the rela-

tionships between the aforementioned concept nodes.

This is in accordance with the ﬁndings of (Sowa,

1976). As outlined in (Ehrlinger and W

oß, 2016),

a KG acquires information and integrates it into an

ontology, subsequently applying a reasoner to derive

new knowledge. Furthermore, in accordance with

the deﬁnition provided by (Ji et al., 2021), KGs are

”structured representations of a fact, consisting of en-

tities, relations, and semantics.” Entities may be either

real-world objects or abstract concepts. Relationships

represent the relationship between entities, and se-

mantic descriptions of entities and their relationships

contain types and properties with deﬁned semantics.

Property graphs, in which nodes and relations possess

properties or attributes, or attribute graphs, are exten-

sively employed. All of these facets rely on a knowl-

edge inference over knowledge graphs, which repre-

sents one of the core technologies in the design of our

ML engineering BoK. The Semantic Web community

has reached a consensus on the use of RDF to repre-

sent a knowledge graph. Then, RDF model also al-

lows for a more expressive semantics of the modeled

data that can be used for knowledge inference. As a

result, a KG is a set of interconnected information on

a speciﬁc set of facts that includes characteristics of

many data management paradigms:

• Database: Structured queries can be used to ex-

plore data in a database.

• Graph: KGs can be analyzed in the same way that

any other network data structure can be.

• Knowledge Base: Formal semantics are encoded

in KGs, which can be used to understand data and

infer new facts.

3.6 Step 5: Knowledge Query

In this context, a body of knowledge (BoK) is con-

ceptualized as a graph of knowledge, as proposed

by (Mattioli et al., 2022). Ultimately, the utility of the

ingested, transformed, integrated and stored knowl-

edge is contingent upon the efﬁciency with which an-

swers can be retrieved by users in an intuitive man-

ner. At the present time, keyword queries and spe-

cialized query languages (e.g. SQL and SPARQL)

represent the prevailing approaches to information re-

trieval. However, in order to facilitate the search for

a speciﬁc ML engineering knowledge by querying

the KG and selecting the set of relevant engineering

views to perform speciﬁc ML engineering activities,

it is necessary to enable the identiﬁcation of simi-

larities between Conﬁance.ai documents by search-

ing for isomorphisms between the graphs represent-

ing the knowledge extracted from the text. A num-

ber of algorithms have been deﬁned which implement

subgraph isomorphism; however, the subgraph iso-

morphic problem is an NP-complete problem. The

initial component is a generic sub-graph matching

mechanism that functions in conjunction with fusion

schemes. This component is responsible for ensur-

ing the structural consistency of the merged informa-

tion with respect to the structures of the initial docu-

ments throughout the fusion process. The fusion ap-

proach is constituted by the similarity and compatibil-

ity functions applied to the members of the graphs to

be fused. The generic fusion algorithm can be adapted

to suit the context in which it is used by adopting these

strategies. The knowledge graph fusion method offers

two additional operations, contingent upon the fusion

strategies employed. Information synthesis is the col-

lection and organization of data on a subject. Infor-

mation is then put together into a network through

information synthesis, where any repetitions are re-

moved. Fusing techniques are used to combine in-

formation about the same thing, even though it is in

different forms. When different sources of informa-

tion are used to create a representation of something,

inconsistencies may appear. This function ﬁnds all

the information in a network that follows a speciﬁc

pattern. The structure of the query graph must match

that of the data graph. To ﬁnd the information query

KMIS 2024 - 16th International Conference on Knowledge Management and Information Systems

336

function, look for a one-to-one mapping between the

query graph and the data graph.

4 ILLUSTRATION ON ML

ROBUSTNESS EVALUATION

The utilization of keyword-based queries has become

a prevalent methodology for enabling non-technical

users to access expansive RDF data sets. At the

present time, the user is able to select an engineering

activity within the graph that utilizes the end-to-end

methodology, and the underlying knowledge will then

be presented to them. The ML Engineering BoK is a

trustworthy ML end-to-end engineering guideline that

engineers should follow throughout their activities. It

is based on a comprehensive set of descriptions, en-

gineering knowledge, metrics, and key performance

indicators, which are capitalized in the BoK. These

elements enable engineers to assess both functional

and non-functional properties. For example, an engi-

neer is seeking information on the assessment of ML

model robustness in the context of the activity of eval-

uating an ML model in order to analyze and charac-

terize the system’s sensitivity to changes in the input,

with a view to determining its overall resilience. For

this engineering activity, the BoK suggests a strategy

made of two successive phases: 1) Robustness test by

sampling and perturbation (empirical evaluation) and

2) Formal veriﬁcation of robustness (formal evalua-

tion). Each step is described by a Capela model and

a textual content. It consists in selecting the most ap-

propriate tool for this robustness test. To make this se-

lection, the key criteria are: the ML Model Algorithm,

the type of data (images, time series, and language),

the type of perturbation and its intensity (examples

of perturbation include image luminance, image blur-

ring, geometric transformation of plane position), and

the target level of robustness. There may be differ-

ent tools for data perturbation and for execution of

the test., or it can be the same tool. With the selected

tool, the speciﬁed perturbation is applied on the Test

Dataset and the ML Model is executed on this per-

turbed dataset and the resulting behavior of the ML

Model is captured.

5 CONCLUSIONS

The objective of building a Trustworthy ML Engi-

neering Knowledge Graph is to facilitate more ef-

fective speciﬁcation, design, comprehension, mon-

itoring, and maintenance of ML-based systems for

system design engineers and ML-based system op-

erations personnel. Ultimately, this should enhance

safety, cyber security, reliability, and performance,

while also improving availability. The construc-

tion of a reliable ML engineering framework en-

tails the utilization of an array of data sources and

artiﬁcial intelligence methodologies, encompassing

knowledge representation, knowledge graphs, seman-

tic networks, high-level information fusion, graph

theory, and numerous other techniques. Conﬁ-

ance.ai’s methodological contributions span the entire

development process of an ML-based system, from

initial speciﬁcation and design through to the com-

missioning and subsequent supervision of operational

deployment, and even to the embedding of the latter

in other systems. These contributions are manifold

and include:

• A taxonomy used in trustworthy AI;

• A complete documentation of the process, in-

cluding modeling of activities and roles, with

elements enabling corporate engineering depart-

ments to implement it;

• A ﬁrst development of a Trustworthy AI ontology,

linking the main concepts of the process and the

taxonomy;

• And a “Body-of-Knowledge” which brings to-

gether all these elements and makes them accessi-

ble on the website of the same name.

Furthermore, the ML-engineering BoK provides sup-

port to stakeholders across the ML value chain, of-

fering invaluable assistance in the elicitation, valida-

tion, and veriﬁcation of safety and cyber-security rel-

evant quality attributes. Based on the Conﬁance.ai

end-to-end methodology, it is able to guarantee that

the heterogeneous requirements of stakeholders are

met, thereby further consolidating its status as a fun-

damental element in the ﬁeld of safety within the con-

text of Machine Learning. While the deployment of

these methodological instruments does not inherently

guarantee the compliance of ML-based systems with

regulatory requirements, it may serve as a basis for

justifying such compliance, particularly in light of the

prevailing standards set by the notiﬁed bodies respon-

sible for veriﬁcation. Furthermore, this ML Engineer-

ing BoK serves as a valuable resource in addressing

the following key challenges:

• How to design AI models, so that, by construc-

tion, they satisfy trustworthy properties (accuracy,

robustness, etc.)?

• How to characterize these AI models, for exam-

ple, to understand and explain their behavior and

their adequacy to the operational domain?

• How to implement and embed those AI models on

hardware, by making them ﬁt for the target with-

ML System Engineering Supported by a Body of Knowledge

337

out losing their trustworthy properties.

• What are the data engineering method to apply in

order to manage important volumes of data, ac-

count for the evolution of the operational domain,

etc.?

• What are the appropriate veriﬁcation, validation,

and certiﬁcation processes to consider for AI-

based systems?

ACKNOWLEDGEMENTS

This work has been supported by the French govern-

ment under the ”France 2030” program, as part of the

SystemX Technological Research Institute within the

Conﬁance.ai Program (www.conﬁance.ai).

REFERENCES

Adedjouma, M., Alix, C., et al. (2022). Engineering de-

pendable ai systems. In 2022 17th Annual System of

Systems Engineering Conference (SOSE), pages 458–

463. IEEE.

Awadid, A., Le Roux, X., et al. (2024). Ensuring the Re-

liability of AI Systems through Methodological Pro-

cesses. In The 24th IEEE International Conference

on Software Quality, Reliability, and Security.

Chen, Z., Wang, Y., et al. (2020). Knowledge graph com-

pletion: A review. IEEE Access, 8:192435–192456.

Ehrlinger, L. and W

oß, W. (2016). Towards a deﬁnition

of knowledge graphs. SEMANTiCS (Posters, Demos,

SuCCESS), 48(1-4):2.

Harnad, S. (1990). The symbol grounding problem. Physica

D: Nonlinear Phenomena, 42(1-3):335–346.

Heist, N., Hertling, S., et al. (2020). Knowledge graphs on

the web–an overview. Knowledge Graphs for eXplain-

able Artiﬁcial Intelligence: Foundations, Applications

and Challenges, pages 3–22.

Hertz, S., Olof-Ors, M., et al. (2019). Machine learning-

based relationship association and related discovery

and search engines. US Patent 10,303,999.

Ji, S., Pan, S., et al. (2021). A survey on knowledge graphs:

Representation, acquisition, and applications. IEEE

transactions on neural networks and learning systems,

33(2):494–514.

Laudy, C. (2011). Semantic knowledge representations for

soft data fusion. Efﬁcient Decision Support Systems-

Practice and Challenges from Current to Future.

Laudy, C., Ganascia, J.-G., and Sedogbo, C. (2007). High-

level fusion based on conceptual graphs. In 2007

10th International Conference on Information Fusion,

pages 1–8. IEEE.

Li, B. et al. (2017). Applications of artiﬁcial intelligence

in intelligent manufacturing: a review. Frontiers

of Information Technology & Electronic Engineering,

18(1):86–96.

Mattioli, J., Laudy, C., et al. (2022). Body-of-knowledge

development by using artiﬁcial intelligence. In 17th

IEEE System of Systems Engineering Conference

(SOSE), pages 14–19.

Mattioli, J., Tachet, D., et al. (2024). Leveraging Knowl-

edge Graph to design the Machine-Learning Engi-

neering Body-of-Knowledge. In IEEE International

Conference on AI x Science, Technology, and Tech-

nology (AIxSET), Laguna hills, United States.

Meyer, L., Stadler, C., et al. (2023). LLM-assisted knowl-

edge graph engineering: Experiments with ChatGPT.

In Working conference on Artiﬁcial Intelligence De-

velopment for a Resilient and Sustainable Tomor-

row, pages 103–115. Springer Fachmedien Wiesbaden

Wiesbaden.

Mugnier, M.-L. and Chein, M. (1992). Conceptual graphs:

Fundamental notions. Revue d’intelligence artiﬁ-

cielle, 6(4):365–406.

Oren, T. (2005). Toward the body of knowledge of model-

ing and simulation. In Interservice/industry training,

simulation, and education conference (I/ITSEC), vol-

ume 2005.

Robert, F., Abran, A., and Bourque, P. (2002). A technical

review of the software construction knowledge area in

the swebok guide. In 10th International Workshop on

Software Technology and Engineering Practice, pages

36–42. IEEE.

Shapiro, S. C. (2006). Knowledge representation. Encyclo-

pedia of cognitive science.

Sowa, J. F. (1976). Conceptual graphs for a data base in-

terface. IBM Journal of Research and Development,

20(4):336–357.

Sowa, J. F. (2000). Guided tour of ontology. Retrieved from.

Strakov

a, J., Straka, M., and Haji

c, J. (2019). Neural ar-

chitectures for nested ner through linearization. arXiv

preprint arXiv:1908.06926.

Wang, C., Liu, X., et al. (2022). Deepstruct: Pretraining

of language models for structure prediction. arXiv

preprint arXiv:2205.10475.

Zeng, X., Zeng, D., et al. (2018). Extracting relational facts

by an end-to-end neural model with copy mechanism.

In Proceedings of the 56th Annual Meeting of the As-

sociation for Computational Linguistics (Volume 1:

Long Papers), pages 506–514.

KMIS 2024 - 16th International Conference on Knowledge Management and Information Systems

338