Knowledge Reinjection Policies and Machine Learning Integration in

CWM-Based Complex Data Warehouses

Fabrice Razaﬁndraibe, Jean Christian Ralaivao, Angelo Raherinirina and Hasina Rakotonirainy

University of Fianarantsoa, Madagascar

Keywords:

Common Warehouse Metamodel (CWM), Knowledge Reinjection, Data Warehouse Architecture, Ontology

Integration, Machine Learning, Semantic Interoperability, Knowledge Management, Metadata

Management, Knowledge Discovery, Ontological Reasoning.

Abstract:

This article discusses the growing complexity of data warehousing systems and the need for enhanced frame-

works that can effectively manage simultaneously metadata and knowledge. While the Common Warehouse

Metamodel (CWM) provides a standardized method for metadata management, its semantic limitations hinder

its use in complex environments. To overcome these shortcomings, the paper proposes an extended CWM

framework that incorporates ontologies, machine learning, and knowledge re-injection policies. This new

framework introduces additional layers and components, such as a ’learning package’ and advanced knowl-

edge mapping, to improve semantic interoperability, adaptability and usability. The research also explores

the integration of hybrid AI systems that use both inductive and deductive methods to facilitate knowledge

discovery and improve decision making.

1 INTRODUCTION AND

RESEARCH QUESTIONS

The increasing complexity of data warehouse projects

requires robust architectures capable of effectively

managing and exploiting the knowledge embedded

in data. Traditional data warehouse systems excel

at handling structured data but struggle to meet the

semantic depth and dynamic requirements of mod-

ern organizations. The Common Warehouse Meta-

model (CWM) provides a standardized framework

for metadata management, promoting interoperabil-

ity and governance. However, its limited semantic

expressiveness poses challenges in representing com-

plex relationships and domain-speciﬁc rules, which

are essential for advanced knowledge integration and

decision-making.

Recent advances in ontology-based modeling

and Artiﬁcial Intelligence (AI), particularly machine

learning, have created new opportunities to improve

knowledge discovery, representation, and utilization

within data warehouses. Ontologies provide a struc-

tured approach to capturing domain knowledge and

semantic relationships, while machine learning tech-

niques facilitate adaptive, data-driven insights. By in-

tegrating these methodologies into CWM-based ar-

chitectures, we aim to address long-standing chal-

lenges in metadata and knowledge management, such

as heterogeneity, scalability, and performance opti-

mization.

This paper addresses the challenge of integrating

knowledge reinjection policies and machine learning

into CWM-based data warehouses to enhance meta-

data management. To achieve this, we explore the fol-

lowing key research questions: 1. How can the CWM

framework be extended to improve semantic expres-

siveness and adaptability? 2. What methodologies

best support the integration of machine learning for

knowledge reinjection and metadata enhancement?

To answer these questions, we propose an ex-

tended CWM framework incorporating ontologies,

machine learning, and systematic knowledge reinjec-

tion policies. This new approach introduces addi-

tional layers and components—such as a Learning

Package and advanced knowledge mapping mecha-

nisms—to enhance semantic interoperability, adapt-

ability, and usability. The research also explores the

integration of hybrid AI systems, combining inductive

and deductive learning methods to facilitate knowl-

edge discovery and improve decision-making.

The rest of the paper is structured as follows:

• Section 2 reviews the state of the art in knowledge

integration, ontology-based systems, and machine

learning applications in data warehouses. • Section 3

describes the proposed methodology and framework,

detailing the CWM extension and knowledge reinjec-

tion policies. • Section 4 presents results.

By addressing these aspects, this work contributes

Razaﬁndraibe, F., Ralaivao, J. C., Raherinirina, A. and Rakotonirainy, H.

Knowledge Reinjection Policies and Machine Learning Integration in CWM-Based Complex Data Warehouses.

DOI: 10.5220/0013352600003929

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

In Proceedings of the 27th International Conference on Enterprise Information Systems (ICEIS 2025) - Volume 1, pages 301-308

ISBN: 978-989-758-749-8; ISSN: 2184-4992

301

to the evolution of knowledge-driven data warehouse

architectures, enhancing their ability to manage com-

plex and heterogeneous data.

2 STATE OF THE ART AND

RELATED WORKS

2.1 Incorporation of Ontologies in Data

Warehouse

Ontology-based modeling has gained signiﬁcant at-

tention as a means to enhance semantic expressive-

ness in data warehouses. Ontologies provide a struc-

tured representation of domain knowledge, enabling

improved metadata management, semantic interoper-

ability, and knowledge reasoning.

Several studies have explored ontology integration

into data warehouses (Falquet et al., 2011; Olsina,

2021). These approaches aim to bridge the gap be-

tween structured data models and semantic knowl-

edge representation. For instance, (Zayakin et al.,

2021) proposed an ontology-based methodology for

analytical platforms in data-intensive domains, em-

phasizing the traceability of metamodel evolution to

maintain semantic relevance. Similarly, (Antunes

et al., 2022) reviewed various ontology classiﬁcations

applicable to data warehouses, highlighting their po-

tential for enhancing data governance and metadata

structuring.

However, despite their beneﬁts, these ontology-

based approaches often lack dynamic knowledge rein-

jection mechanisms. They primarily focus on static

ontology integration, without addressing how newly

acquired knowledge can be reinjected into metadata

management systems dynamically. This limitation

reduces their adaptability in environments requiring

continuous knowledge updates.

2.2 Combining Machine Learning and

Ontology

The integration of machine learning (ML) and on-

tologies is a growing area of research, aimed at en-

hancing knowledge discovery and automated reason-

ing. (Ghidalia et al., 2024) conducted a system-

atic review of hybrid AI systems combining induc-

tive learning (ML-based inference) with deductive

reasoning (ontology-driven approaches). Their work

categorizes hybrid AI applications into: • Learn-

ing-Enhanced Ontologies: Using ML to automate

ontology construction and maintenance. • Seman-

tic Data Mining: Leveraging ontological knowledge

to enhance ML-based data mining. • Learning and

Reasoning Systems: Combining symbolic reasoning

with ML to emulate human-like cognitive processes.

These studies demonstrate the potential of hybrid

AI approaches for adaptive knowledge management.

However, they do not provide a formalized method

for metadata management in CWM-based architec-

tures. Most existing works focus on either ML-driven

knowledge extraction or ontology-based reasoning,

but lack a comprehensive framework for continuous

knowledge reinjection and metadata evolution within

data warehouses.

2.3 CWM Extensions

The CWM is a widely adopted standard for metadata

management in data warehouses, providing a struc-

tured framework for interoperability, governance, and

data lineage tracking. While CWM offers a robust

foundation, its limited semantic expressiveness hin-

ders its application in complex knowledge-driven en-

vironments.

Several studies have proposed CWM extensions

to enhance knowledge discovery and management

(Gomes et al., ; Midouni et al., ; da Silva et al., ; Tavac

and Tavac, ). Among them, (Thavornun, 2015) intro-

duced an RDF-based metadata management approach

for knowledge discovery, extending the Core (Object

layer) and Transformation and Data Mining (Analysis

layer) packages with three key classes: Evidence –

Extracted knowledge patterns; UserAction – User in-

teractions for adaptive learning; DataProperty – Re-

lationships between extracted knowledge elements.

Recent research has explored CWM extensions

to incorporate ontology-based capabilities. For in-

stance, (Ralaivao et al., 2024) proposed extending

CWM with a new knowledge layer (2), introducing:

1. Acquired and Explicit Knowledge – Capturing

technical knowledge and operational statistics. 2. Do-

main Knowledge and Ontology – Structuring spe-

cialized domain-speciﬁc metadata. 3. Metadata and

Knowledge Mapping – Managing transformations

and semantic linkages within CWM.

(Ralaivao et al., 2024) proposed Core package ex-

tension to enable compatibility with the ODM meta-

model. The addition of the class AnnotatedElement

is proposed, would inherit directly from ModelEle-

ment and serves as the parent class for Ontology

and OntologyElement, as depicted in Figure 1. This

proposed extension aims to enhance the CWM meta-

model’s capacity for knowledge representation and

integration, particularly in data-intensive domains.

While these efforts enhance semantic metadata

representation, they do not fully address the integra-

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

302

Figure 1: Elements structuring the ODM metamodel and extending Core Package.

tion of dynamic knowledge reinjection policies.

Figure 2: New layer integrated in CWM metamodel.

3 METHODOLOGY AND

CONTRIBUTION

3.1 Research Methodology

This research addresses knowledge integration, ontol-

ogy enhancement, and machine learning incorpora-

tion into CWM. While CWM supports classical data

mining, it lacks dynamic knowledge reinjection and

adaptive metadata management. Our approach fol-

lows three key steps:1. Extending CWM – Deﬁn-

ing a structured knowledge reinjection policy to en-

sure continuous metadata evolution (Ralaivao et al.,

2024). 2. Integrating AI-driven Metadata Manage-

ment – Leveraging ontology-based machine learn-

ing for adaptive semantic enrichment. 3. Evaluating

the Framework – Assessing performance through

benchmarks and real-world use cases. Key challenges

include metamodel complexity, heterogeneous data

integration, and business-aligned reinjection policies.

Future research should reﬁne automated knowledge

discovery and reinjection to enhance adaptability.

3.2 Contribution of this Paper

This paper extends CWM with a structured knowl-

edge reinjection process and AI-driven enhance-

ments, addressing limitations in existing ontology and

machine learning-based approaches. • AI-Enhanced

CWM – Introducing a Learning Package for adap-

tive metadata enrichment. • Structured Knowledge

Reinjection – Ensuring continuous, validated knowl-

edge updates. • Optimized Query Processing –

Leveraging reinjected knowledge to improve seman-

tic search and decision support.

This hybrid AI and ontology-based framework en-

ables scalable, real-time metadata management with

enhanced semantic expressiveness.

4 RESULTS

4.1 Knowledge Reinjection Policy

4.1.1 General Reinjection Framework

Unlike the traditional approach of simply extracting

knowledge from data, this research aims to take the

knowledge derived from the data, reclassify it, vali-

date it and then recode it in the form of actionable

constraints. These constraints are intended to be used

at the data level, particularly in the form of enriched

metadata, to optimise data management and decision-

making processes.

Knowledge Reinjection Policies and Machine Learning Integration in CWM-Based Complex Data Warehouses

303

For instance, in the healthcare domain, the reinjec-

tion process detects inconsistencies in patient data and

updates metadata dynamically. Suppose an AI model

identiﬁes a new correlation between symptoms and

diagnoses. The system classiﬁes this as new knowl-

edge, validates it with existing ontologies, and rein-

jects it into the metadata layer for future queries.

The proposed knowledge reinjection process in-

volves several key strategies:

1. Optimization of Query Structure The charac-

teristics of a variable—such as its probability distribu-

tion and parameters—can enhance query performance

in a data warehouse. By leveraging these characteris-

tics, you can reﬁne the order in which predicates are

applied within query selections. This strategic appli-

cation can lead to signiﬁcant improvements in query

efﬁciency; 2. Uncovering Hidden Functional De-

pendencies The process also involves the discovery

of non-explicit functional dependencies within large

volumes of data. These dependencies, often over-

looked, can be leveraged to optimize the design and

organization of the data warehouse schema, ensuring

a more efﬁcient and effective structure.; 3. Enrich-

ing Knowledge and Data Structures The validated

knowledge, once reﬁned, becomes accessible to Com-

plex Data Warehouse administrator. This knowledge

can be used to: • Enrich the Ontology: Identify and

add new concepts, uncover relationships between ex-

isting concepts, and challenge or reﬁne existing on-

tological links; • Reformulate Queries: Use newly

discovered knowledge to revise queries, incorporat-

ing newly identiﬁed concepts, more relevant calcula-

tion rules, and optimized query formulations; • Reor-

ganize Complex Data Warehouse Structures: In-

troduce new hierarchies or reorganize schema di-

mensions, update CWM’s ”Business Nomenclature”

layer, or develop new transformation and deduction

rules to reﬂect new insights; • Adjust Administra-

tion Parameters: Optimize the administration of the

Complex Data Warehouse, including adjusting ETL

scheduling, process management, and aligning pro-

cesses with the ”Warehouse Process” and ”Warehouse

Operation” layers within CWM, ensuring smoother

and more effective operations.

Once the knowledge has been validated, it is an-

notated and supported by the “Knowledge” layers of

the extended CWM, in particular the “MD & K Map-

ping” package. Depending on the annotation attached

to each piece of knowledge, the latter is responsible

for transcribing it and distributing it to the various

layers concerned (as new knowledge, obsolete knowl-

edge or modiﬁed knowledge).

4.1.2 Reinjection Process

The knowledge reinjection process (Figure 3) fol-

lows a structured workﬂow to ensure that new knowl-

edge is effectively classiﬁed, validated, and integrated

into the system. It consists of the following steps:

1. Knowledge Extraction: Identify new patterns

from the data warehouse using machine learning tech-

niques. 2. Validation: Compare extracted knowl-

edge with existing ontologies to ensure consistency.

3. Metadata Update: Integrate new insights into the

metadata management layer. 4. Query Optimiza-

tion: Adapt the query engine to leverage enriched

metadata.

Each new knowledge entry is categorized as either

conﬁrmed, tentative, or rejected, ensuring robust con-

trol over metadata evolution.

Figure 3: Knowledge reinjection procedure.

If it is decided to transform the knowledge into

metadata, the ‘MD & K’ layer is responsible for this

transformation and for injecting it into the layer(s)

concerned. If it is decided to keep it in the form of

knowledge, the same ‘MD & K’ layer is responsible

for transmitting it to the ‘Ontology’ layer. This in turn

decides whether to update the Tbox or Abox bases of

the ontology in DL. This then updates the layers con-

cerned in the same way as the metadata. Note that in

this second case, it is mainly the new layers resulting

from the CWM extension that are most involved.

4.2 CWM Learning Integration

We propose to take the classiﬁcations deﬁned by

(Ghidalia et al., 2024) and transform them into the

concept of classes for possible integration into the

CWM metamodel. And we propose to extend the

Analysis layer by adding a Learning package (Figure

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

304

4). This new package will work in conjunction with

the Knowledge layer to improve knowledge manage-

ment and governance.

To further clarify the role of machine learning

within our extended CWM framework, we categorize

its functions into three main strategies:

Table 1: ML Strategies for Knowledge Reinjection.

Strategy Objective Algorithms

Knowledge

Discovery

Extract patterns

and new insights

from raw data.

Decision Trees,

Random Forest,

Autoencoders

Ontology

Enrichment

Automatically

update and reﬁne

ontology struc-

tures.

Word2Vec for

semantic similar-

ity, Graph Neural

Networks

Reinjection

Optimiza-

tion

Decide when and

how to reinject

new knowledge.

Reinforcement

Learning, Bayesian

Networks

By formalizing these strategies, our framework

ensures that machine learning contributes effectively

to metadata enhancement and knowledge manage-

ment.

4.2.1 The MLModel Component

MLModel represents the abstract concept unifying

the three components and uses the AnnotedElement

concept from extended Core Package. It deﬁnes

common attributes like reasoningType and knowl-

edgeBase and methods like dataIntegration(), train-

Model(), and infer().

Properties are characterised as follows: 1. rea-

soningType inﬂuences how knowledge is processed

to derive new insights or verify existing information.

It shapes the system’s reasoning strategy, tailored to

its speciﬁc application domain and desired outcomes.

Common reasoning types can be : Deductive, In-

ductive, Abductive, Ontological or Hybrid reasoning.

2. knowledgeBase serves as the factual and logical

foundation that enables the AI system to comprehend

its domain and reach informed conclusions. It ensures

the system has access to reliable and organized infor-

mation. The components of a KnowledgeBase can be

: T Box, ABox, Logical Rules or Data Sources.

The methods are designed as follows: 1. dataIn-

tegration() method uniﬁes and prepares disparate

data sources to create a cohesive dataset for analy-

sis and learning. It links raw data with ontologi-

cal knowledge, ensuring semantic alignment and data

consistency. 2. trainModel() method trains a ma-

chine learning model using integrated and pre-pro-

cessed data. It uses domain-speciﬁc knowledge from

ontologies to enhance the training process, ensuring

that the model effectively learns patterns and relation-

ships relevant to the target domain. 3. infer() method

applies the trained model to new data, performing in-

ference tasks by integrating knowledge from both the

ontology and the learned model. It generates predic-

tions and explanations based on data patterns and on-

tological rules.

Classes like LearningEnhancedOntology,

SemanticDataMining, and LearningAndReason-

ingSystem are abstracted under MLModel, with

relevant relationships to supervised, unsupervised,

and neural techniques.

4.2.2 The LearningEnhancedOntology

Component

The LearningEnhancedOntology component auto-

mates the extraction and classiﬁcation of ontological

concepts. For instance, in an e-commerce system, this

module can dynamically classify new product cate-

gories based on customer interactions and reinject this

knowledge into the search engine’s metadata to en-

hance recommendations. It includes attributes like

ontologyType and ruleset.

1. ontologyType deﬁnes the speciﬁc type of ontol-

ogy used in the system. Examples include domain on-

tology, task ontology and parent ontology. This clas-

siﬁcation helps the system to understand the focus and

scope of the ontology and facilitates the adaptation

of learning processes to the type of ontology. In the

education domain, the ontology type could be called

”curriculum ontology”, which includes courses, pre-

requisites and learning objectives. 2. ruleset deﬁnes

the logical rules and constraints that govern the on-

tology. It provides a deductive framework for reason-

ing and knowledge inference, while ensuring compli-

ance with domain-speciﬁc rules. 3. conceptualMap-

ping indicates whether an ontology supports connec-

tions between various knowledge structures, such as

schemas or ontologies. This capability facilitates in-

teroperability and data integration, allowing for se-

mantic alignment across multiple data sources and

systems. In a multilingual educational platform, con-

ceptual mapping links equivalent concepts across dif-

ferent language-speciﬁc ontologies.

1. ontologyCreation() method automates the de-

velopment of new ontologies by extracting concepts,

relationships and properties from existing data. This

streamlines the design process and minimises man-

ual effort, ensuring that the resulting ontologies are

data-driven and accurately represent the domain. It

can extract data from both structured and unstruc-

tured sources, identifying key concepts and relation-

ships using machine learning algorithms such as clus-

Knowledge Reinjection Policies and Machine Learning Integration in CWM-Based Complex Data Warehouses

305

Figure 4: Extensible Learning Package for extending CWM Analysis layer.

tering and neural networks. The extracted knowledge

is then formalised in ontology formats such as OWL

or RDF. 2. ontologyMaintenance() function updates

and reﬁnes the ontology to ensure its relevance and

accuracy over time. It keeps the ontology synchro-

nised with changes in the domain or data, support-

ing the dynamic evolution of knowledge structures. It

can monitor domain changes or new data input and

automatically add, modify or delete concepts and re-

lationships. Updates are validated using domain-spe-

ciﬁc rules or expert feedback. 3. optimizeReason-

ing improves the reasoning process by incorporating

machine learning techniques, resulting in faster and

more accurate reasoning. It also reduces the compu-

tational complexity of ontology-based reasoning, al-

lowing reasoning engines to adapt to real-time appli-

cations. The processes can be : • Identify bottle-

necks: Recognise the limitations of traditional rea-

soning methods, such as long inference times. • Ap-

ply Machine Learning Models: Use models such

as Recursive Reasoning Networks or heuristic opti-

mizations to increase efﬁciency. • Validate results:

Ensure accuracy by comparing results against estab-

lished ontological rules. This approach can improve

the system’s ability to predict the optimal course se-

quence for students based on their learning history.

4. mapOntologies() effortlessly aligns disparate on-

tologies, paving the way for seamless data integration

and interoperability. It fuses knowledge from differ-

ent sources to present a uniﬁed view, while resolving

any annoying semantic inconsistencies. First things

ﬁrst: identify common or equivalent concepts across

ontologies. We use sophisticated algorithms such as

Random Forest or Word2Vec to align concepts and

their relationships. Once we’ve mapped them, we val-

idate the relationships to ensure everything is consis-

tent and accurate.

4.2.3 The SemanticDataMining Component

SemanticDataMining ocuses on feature engineer-

ing, algorithm design, domain knowledge integration,

and model explanation. It includes attributes like

dataRepresentation and embeddingMethod.

1. dataRepresentation deﬁnes how data is struc-

tured before it enters the machine learning pipeline.

It ensures that the data is compatible with both the

data mining algorithms and the associated ontologi-

cal framework. This process determines whether to

use raw data, semantic embeddings or ontologically

annotated data. In short, effective data representation

is critical to harnessing the power of machine learning

technologies. A use case could be ”Learning Activity

Features with Ontological Annotations”. 2. embed-

dingMethod transforms data into a format rich in se-

mantic knowledge. This process enables data mining

algorithms to effectively utilize ontological informa-

tion. By integrating these techniques, we can enhance

the capabilities of our data analysis and improve de-

cision-making processes. We propose ”Word2Vec for

Learning Object Metadata” as an example. 3. seman-

ticConstraints refers to the application of semantic

rules during the data mining process. This feature en-

sures that machine learning models follow the speciﬁc

rules and relationships outlined by the underlying on-

tology, thereby maintaining domain relevance and in-

tegrity.

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

306

1. featureEngineering() function enhances, se-

lects or extracts features from raw data by incor-

porating ontological knowledge. This process im-

proves the relevance and representation of data for

the machine learning pipeline. The types of Feature

Engineering are : a) Feature Augmentation: Adds

additional semantic features derived from the ontol-

ogy b) Feature Selection: Identiﬁes the most rele-

vant features based on ontological rules c) Feature

Extraction: Transform raw data into meaningful fea-

tures using ontology-based embeddings. The fea-

ture engineering process includes the following steps

: a) Extract raw features from the dataset. b) Us-

ing the ontology to add, reduce or transform fea-

tures. c) Outputting a semantically enriched feature

set. By following these steps, we can create a more

effective and informative dataset for machine learn-

ing applications. 2. designAlgorithm() function de-

signs or adapts data mining algorithms to incorpo-

rate ontological knowledge, aligning the logic and

behaviour of the algorithm with domain-speciﬁc se-

mantics. Types of Ontology-Enhanced Algorithms

are like a) Ontology-based decision trees: These use

decision nodes that exploit semantic relationships.

b) Ontology-based neural networks: The layers or

architecture of these networks are inﬂuenced by on-

tological hierarchies. c) Ontology-based probabilis-

tic graphical models: In these models, relationships

are driven by ontology constraints. Identifying the

data mining problem (e.g. classiﬁcation or cluster-

ing). Adapting an existing algorithm or creating a

new one that incorporates ontology-driven rules. Val-

idating the performance of the algorithm on semanti-

cally enriched data sets. 3. integrateDomainKnowl-

edge() incorporates domain-speciﬁc knowledge from

an ontology into the data mining pipeline. This ap-

proach ensures that machine learning models beneﬁt

from structured and contextual information. Steps to

Implement: a) Load Domain Ontology: Import the

relevant domain ontology associated with the dataset.

b) Identify Key Concepts: Recognize essential con-

cepts, relationships, and constraints that need to be

integrated. c) Apply Ontological Insights: Utilize

the identiﬁed insights to inform preprocessing, model

training, and validation. This structured approach

enhances the machine learning process by ground-

ing it in relevant domain knowledge. 4. explain-

Model() improves model interpretability by linking

predictions or decisions to ontological concepts and

relationships. It increases the transparency and trust-

worthiness of machine learning models by providing

domain-relevant justiﬁcations for model outputs. The

steps are to analyse the model’s predictions or deci-

sion paths, map these outputs to relevant ontologi-

cal concepts and rules, and generate explanations that

align the predictions with domain knowledge.

4.2.4 The LearningReasoningSystem Component

The LearningReasoningSystem leverages hybrid AI

techniques to dynamically adjust its reasoning rules.

For example, in a ﬁnancial risk assessment system,

it can detect fraudulent patterns by cross-referencing

new transactional data with historical fraud cases,

thereby reﬁning its predictive accuracy in real-time.

It combines learning and reasoning capabilities

with scalability and neuro-symbolic integration. It in-

cludes attributes like modelArchitecture and methods

like dynamicAdaptation() and realTimeReasoning().

The properties are crucial for the efﬁciency and

adaptability of hybrid AI. These properties include:

1. modelArchitecture: This deﬁnes the structure of

the learning model used within the system. It could be

neural networks, decision trees, or a hybrid approach.

The modelArchitecture directly impacts how the sys-

tem processes data and performs reasoning, ensur-

ing it’s well-suited for the application’s complexity.

For example, in a dynamic recommendation system,

a ”Recurrent Neural Network with Ontology-Based

Constraints” architecture could optimize suggestions

by considering semantic relationships between items.

2. reasoningMechanism: This speciﬁes the reason-

ing framework used to infer knowledge or validate

decisions. This could involve rule-based, probabilis-

tic, or hybrid reasoning. The reasoningMechanism

guides decision-making by leveraging domain-spe-

ciﬁc logic and combining data-driven learning with

deductive reasoning. For instance, in a ﬁnancial fraud

detection system, a ”Probabilistic Reasoning with

Bayesian Networks” mechanism could evaluate sus-

picious transactions by considering the probabilities

of fraudulent behavior. 3. dynamicKnowledgeBase:

This indicates whether the system’s knowledge base

is dynamically updated with new information. A dy-

namicKnowledgeBase allows the system to adapt to

evolving domains and learn in real-time. A medi-

cal diagnostic system that continuously updates its

knowledge base with the latest research and treat-

ment recommendations is a good example, ensuring

more precise and relevant analyses. These proper-

ties a powerful and adaptable tool. It can evolve with

changing contexts, improve the quality of decisions,

and provide a deeper understanding of the domains it

models.

The LearningReasoningSystem combines learn-

ing models with reasoning engines to improve

decision-making, adaptability, and scalability. It re-

lies on four key methods:

1. buildDynamicKnowledge() integrates machine

Knowledge Reinjection Policies and Machine Learning Integration in CWM-Based Complex Data Warehouses

307

learning with reasoning engines to create a hybrid

system. Machine learning detects patterns, while rea-

soning engines reﬁne predictions using rules or ontol-

ogy-based logic. This improves accuracy and explain-

ability. For example, in a tutoring system, it predicts

student knowledge gaps and adjusts recommenda-

tions based on prerequisite rules. 2. buildDynamic-

Knowledge() keeps the knowledge base up to date by

integrating new data and insights. The system contin-

uously monitors input, updates validated information,

and ensures consistency. In healthcare, it adds new

treatments and diseases dynamically, ensuring diag-

nostic models stay current. 3. scaleReasoning() en-

hances the system’s ability to process large and com-

plex datasets efﬁciently. Optimization techniques,

such as heuristic algorithms and parallel computing,

improve reasoning performance. In urban planning,

this method analyzes trafﬁc data from sensor net-

works to suggest optimal routes. 4. integrateNeu-

roSymbolicApproaches() combines neural networks

with symbolic reasoning for better decision-making.

Neural networks handle unstructured data, while sym-

bolic reasoning reﬁnes predictions for more reliable

results. In legal document analysis, it classiﬁes docu-

ments with AI and validates classiﬁcations using le-

gal rules. These methods make the LearningRea-

soningSystem more adaptive, scalable, and capable

of delivering precise and interpretable decisions.

5 CONCLUSION

The CWM framework offers a standardized approach

to metadata management, yet its limited semantic

expressiveness restricts its applicability in complex,

heterogeneous environments. This paper proposes

an extended CWM framework that incorporates on-

tologies, machine learning, and knowledge reinjec-

tion policies to address these limitations. By in-

troducing new layers and components, such as the

”Learning Package” and advanced knowledge map-

ping mechanisms, the framework enhances the se-

mantic interoperability, adaptability, and usability of

data warehouse systems. The research explores the

integration of hybrid AI systems, combining induc-

tive and deductive techniques, to improve knowl-

edge discovery and decision-making. With the ex-

tension of the CWM mentioned in (Ralaivao et al.,

2024) and the present extension, we obtain an ex-

tended CWM framework capable of discovering and

managing knowledge. As a way forward, we pro-

pose experimentation and benchmarking on simple

and knowledge-based queries. Future work will fo-

cus on reﬁning the reinjection mechanism with adap-

tive learning techniques and evaluating its applica-

bility across various domains, such as education and

learning, healthcare, ﬁnance, and smart cities.

REFERENCES

Antunes, A. L., Cardoso, E., and Barateiro, J. (2022). In-

corporation of ontologies in data warehouse/business

intelligence systems-a systematic literature review.

2(2):100131.

da Silva, J., de Oliveira, A. G., do Nascimento Fidalgo,

R., Salgado, A. C., and Times, V. C. Modelling and

querying geographical data warehouses. 35:592–614.

Falquet, G., M

etral, C., Teller, J., Tweed, C., Roussey, C.,

Pinet, F., Kang, M. A., and Corcho, O. (2011). An

introduction to ontologies and ontology engineering.

Ontologies in Urban development projects, pages 9–

38.

Ghidalia, S., Narsis, O. L., Bertaux, A., and Nicolle,

C. (2024). Combining machine learning and ontol-

ogy: A systematic literature review. arXiv preprint

arXiv:2401.07744.

Gomes, P., Farinha, J. T., and Trigueiros, M. J. A data qual-

ity metamodel extension to cwm. In Asia-Paciﬁc Con-

ference on Conceptual Modelling.

Midouni, S. A. D., Darmont, J., and Bentayeb, F. Approche

de mod

elisation multidimensionnelle des donn

ees

complexes : application aux donn

ees m

edicales,. In

Journ

ees Francophones sur les Entrep

ots de Donn

ees

et l’Analyse en ligne.

Olsina, L. (2021). The foundational ontology thingfo: Ar-

chitectural aspects, concepts, and applicability. In In-

ternational Joint Conference on Knowledge Discov-

ery, Knowledge Engineering, and Knowledge Man-

agement, pages 73–99. Springer.

Ralaivao, C., Razaﬁndraibe, F., Raherinirina, A., and Rako-

tonirainy, H. (2024). Cwm extensions for knowl-

edge and metadata integration for complex data ware-

house and big data. In Proceedings of the 26th

International Conference on Enterprise Information

Systems, Angers, France, volume 1, pages 329–336.

SCITEPRESS - Science and Technology Publications.

Tavac, M. and Tavac, V. The general algorithm for the de-

sign of the mda transformation models. pages 171–

176.

Thavornun, V. (2015). Metadata management for knowl-

edge discovery. Master’s thesis.

Zayakin, V. S., Lyadova, L. N., Lanin, V. V., Zamyatina,

E. B., and Rabchevskiy, E. A. (2021). An ontology-

driven approach to the analytical platform develop-

ment for data-intensive domains. In International

Joint Conference on Knowledge Discovery, Knowl-

edge Engineering, and Knowledge Management, vol-

ume 1718, pages 129–149. Springer, Springer Inter-

national Publishing.

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

308