IoT-AID: Leveraging XAI for Conversational Recommendations in

Cyber-Physical Systems

Mohammad Choaib

1,2

, Moncef Garouani

, Mourad Bouneffa

and Adeel Ahmad

Univ. Littoral C

ote d’Opale, UR 4491, LISIC, Laboratoire d’Informatique Signal et Image de la C

ote d’Opale, F-62100

Calais, France

Lebanese University, Lebanon

IRIT, UMR 5505 CNRS, Universit

e Toulouse Capitole, Toulouse, France

Keywords:

Cyber Physical System, Conversational Recommender Systems, Machine Learning, Industry 4.0, Decision

Support Systems, NLP, Large Language Models, Explainable AI.

Abstract:

The rapid evolution of Industry 4.0 has introduced transformative technologies such as the Internet of Things

(IoT), Artiﬁcial Intelligence (AI), and big data, facilitating real-time data collection, processing, and decision-

making. At the heart of this revolution lies Cyber-Physical Systems (CPS), which integrate computational

algorithms with physical components to create intelligent, resilient, and adaptive systems. However, CPS

deployment remains complex due to the need for extensive domain expertise. This paper introduces IoT-

AID, a novel Explainable AI (XAI)-driven Cyber-Physical Recommendation System (CPRS) that enhances

transparency, trust, and efﬁciency in CPS design. IoT-AID integrates traditional machine learning models,

deep learning architectures, and ﬁne-tuned transformer-based models with XAI techniques to automate and

improve CPS conﬁguration. Our approach ensures that AI-driven recommendations are interpretable, thereby

increasing adoption across industries.

1 INTRODUCTION

Industry 4.0 marks a pivotal era where the fusion of

IoT, AI, and big data transforms industries, enabling

real-time monitoring, predictive maintenance, and de-

centralized decision-making (Choaib et al., 2024).

Cyber-Physical Systems (CPS) form the backbone of

this transformation, integrating physical components

with computational intelligence to optimize industrial

processes. Despite their potential, CPS deployment

faces key challenges. One of the major issues is

the complexity of conﬁguring such systems, as en-

gineers must integrate multiple components, requir-

ing deep domain expertise in hardware and software

(Garouani et al., 2022a). Additionally, the availabil-

ity of high-quality domain-speciﬁc datasets, remain a

bottleneck for training AI models effectively (Whang

et al., 2023). Many CPS recommendation systems

also lack transparency, often functioning as black-box

models, making it difﬁcult for users to trust their out-

puts and interpret how recommendations are gener-

ated (Garouani et al., 2022c).

To address these challenges, IoT-AID has been de-

veloped as a novel solution that integrates Explain-

able AI (XAI) techniques with advanced AI mod-

els, ensuring transparent and user-centric CPS design

recommendations (Moosavi et al., 2024). By imple-

menting a multi-faceted approach combining machine

learning, deep learning, and transformer-based mod-

els, IoT-AID enhances the accuracy and interpretabil-

ity of recommendations while simplifying the conﬁg-

uration process for engineers and decision-makers.

Building effective CPS applications necessitates

a deep understanding of user needs and the accurate

identiﬁcation of necessary physical components, such

as sensors. While Large Language Models (LLMs)

have revolutionized various aspects of Industry 4.0,

their potential in smart manufacturing and Cyber-

Physical Systems (CPS) remains largely unexplored

(Choaib et al., 2024). LLMs, which are trained on

extensive, generalized knowledge, often lack the spe-

cialized insights required to navigate the complex

challenges of these domains (Zaheer et al., 2020). To

bridge this gap, we propose ﬁne-tuning LLMs with

datasets speciﬁcally built for entity recognition, en-

hancing their ability to generate customized recom-

Choaib, M., Garouani, M., Bouneffa, M. and Ahmad, A.

IoT-AID: Leveraging XAI for Conversational Recommendations in Cyber-Physical Systems.

DOI: 10.5220/0013497100003929

In Proceedings of the 27th International Conference on Enterprise Information Systems (ICEIS 2025) - Volume 1, pages 671-679

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

671

mendations for CPS conﬁguration. However, we also

leveraged traditional machine learning models such

as Decision Trees, Support Vector Machines (SVM),

and Naive Bayes, as well as deep learning models

like Recurrent Neural Networks (RNN) and Convo-

lutional Neural Networks (CNN). This multi-faced

approach, combined with ﬁne-tuning a pre-trained

BERT model for text classiﬁcation, ensured a com-

prehensive understanding of user needs and facilitated

the recommendation of essential components. By

integrating these techniques into the Cyber-Physical

Recommender System (CPRS), we automated and en-

hanced the recommendation process, leading to more

efﬁcient CPS design and implementation in the realm

of Industry 4.0.

The rest of the paper is organized as follows, a

small background on text classiﬁcation and its differ-

ent models, in addition to the state of the art for cy-

ber physical systems and utilizing XAI. Section 3 dis-

cusses the main components of the proposed system

and we also discuss how these components collabo-

rate to achieve the pursued goals, and how to utilize

different models in this system, Finally in section 4

we will discuss the obtained results and what are the

faced challenges and concludes the paper and outlines

future perspectives in section 5.

2 BACKGROUND

2.1 Text Classiﬁcation

Text classiﬁcation, also referred to as text catego-

rization, assigns predeﬁned labels to text based on

its content, playing a crucial role in various NLP

applications such as sentiment analysis, spam de-

tection, topic categorization, and document sum-

marization. Traditionally, methods for text classi-

ﬁcation ranged from simple rule-based systems to

more complex machine learning models (Sebastiani,

2002). However, recent advancements in deep learn-

ing and Transformer-based architectures have signiﬁ-

cantly enhanced performance and expanded the range

of applications (Devlin et al., 2019; Yang et al., 2019).

This machine learning subﬁeld, akin to recognizing

features of different ﬂowers, involves training algo-

rithms to detect patterns in words and phrases, en-

abling them to classify new, unlabeled texts. This

technique is utilized in a variety of contexts, including

ﬁltering spam emails (Mardiansyah and Surya, 2024),

analyzing social media sentiment for hate speech

(Zampieri et al., 2023), and categorizing news articles

and videos by topic (Zaheer et al., 2020), highlighting

its versatile and practical applications.

2.1.1 Traditional Machine Learning Methods

• Naive Bayes: A probabilistic classiﬁer based

on Bayes’ theorem, assuming independence be-

tween features. It is simple, fast, and effective

for many text classiﬁcation tasks but can struggle

with highly correlated features (McCallum and

Nigam, 1998).

• Support Vector Machines (SVM): A discrimi-

native classiﬁer that ﬁnds a hyperplane to sepa-

rate data points of different classes. Known for its

effectiveness in high-dimensional spaces, SVMs

provide strong performance for text classiﬁcation

tasks (Joachims, 1998).

• Decision Tree: A ﬂowchart-like model where

each node represents a feature and each branch

represents a decision rule. While easy to interpret,

decision trees can overﬁt the training data without

proper pruning (Quinlan, 1986).

2.1.2 Deep Learning Methods

• Feedforward Neural Networks (FNN): Simple

neural networks with one or more hidden layers.

While foundational, they are less effective for text

due to the lack of sequential data handling capa-

bilities (Goodfellow et al., 2016).

• Convolutional Neural Networks (CNN): Utilize

convolutional layers to capture local patterns in

text. They are particularly effective for sentence

classiﬁcation and sentiment analysis (Zhao and

Wu, 2016; Garouani et al., 2023).

• Recurrent Neural Networks (RNN) and Long

Short-Term Memory (LSTM): Capture sequen-

tial dependencies in text data, making them suit-

able for text classiﬁcation where context is im-

portant. LSTMs, in particular, address the van-

ishing gradient problem of RNNs (Hochreiter and

Schmidhuber, 1997).

2.1.3 Transformer-Based Models

• BERT (Bidirectional Encoder Representations

from Transformers): Uses transformers to cap-

ture context from both directions in text, setting a

new state-of-the-art for many NLP tasks, includ-

ing text classiﬁcation (Devlin et al., 2019).

• GPT (Generative Pre-trained Transformer): A

generative model that predicts the next word in a

sequence, useful for both text generation and clas-

siﬁcation tasks. GPT-3, in particular, has shown

remarkable capabilities (Brown et al., 2020).

• XLNet: Combines the strengths of autoregressive

and autoencoding models, outperforming BERT

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

672

Natural Language

Understanding

Response Generation

+Explainability of

Recommendation

NLP Module

Preprocessing

Module

Entity

Recognition

Data Analysis

Module

ML models/ Rule

Based models

Semantic analysis

based on

RNN/LSTM

Recommender

Module

Entity

Recognition

DataSet

Response

Generation

Predict Top N

Recommendations

Figure 1: CPRS Architecture.

on various NLP benchmarks by capturing bidi-

rectional contexts without masking (Yang et al.,

2019).

2.2 Cyber-Physical Systems

Cyber-Physical Systems (CPS) integrate computa-

tional algorithms with physical elements to create in-

terconnected systems that offer advanced functional-

ities and capabilities. CPSs leverage smart sensors,

embedded systems, cloud computing, data storage,

and artiﬁcial intelligence techniques to transform in-

dustries, paving the way for smarter factories that are

at the forefront of the fourth industrial revolution.

These advancements enable predictive maintenance,

real-time monitoring, and self-optimization(Garouani

et al., 2022b). As critical components of this rev-

olution, CPSs contribute to the development of in-

telligent, resilient, and adaptive machines, facilitat-

ing their widespread adoption across various sectors

and applications. Despite their potential, conﬁguring

CPS solutions to meet speciﬁc needs remains chal-

lenging for researchers and engineers due to knowl-

edge gaps. Automated assistance can help by en-

abling engineers and researchers to efﬁciently de-

velop, validate, and deploy CPS solutions, thereby

enhancing service quality, productivity, and reducing

dependence on human expertise .

The proposed Cyber-Physical Recommender Sys-

tem (CPRS), illustrated in Figure 1, comprises a

Knowledge Base (KB) and a recommendation engine.

The KB contains comprehensive information about

sensors, including their speciﬁcations and application

domains. Users interact with the system through a

chatbot, providing details such as product informa-

tion and budget constraints. To address new chal-

lenges, the system employs Natural Language Pro-

cessing (NLP) to extract keywords and match them

with data stored in the KB . Recommendations are re-

ﬁned based on rankings and user feedback. Further-

more, a Meta-knowledge base stores knowledge ac-

quired during ofﬂine training and uses ontologies to

enhance information retrieval and query understand-

ing (Choaib et al., 2024).

To improve decision-making, the CPRS incorpo-

rates LLMs such as BERT and some other deep learn-

ing models such as CNN and LSTM. Although LLMs

are proﬁcient in general language understanding, they

may lack the domain-speciﬁc knowledge needed for

cyber-physical systems (Devlin et al., 2019). To miti-

gate this, ﬁne-tuning is performed, training LLMs on

domain-speciﬁc datasets to understand the terminol-

ogy and context of cyber-physical systems (Choaib

et al., 2024). The CPRS uses LLMs to enhance

Natural Language Understanding (NLU), extract en-

tities from user inputs, generate recommendations,

and continuously optimize based on user feedback.

Evaluation metrics include the accuracy of entity ex-

traction and the relevance of recommendations, ulti-

mately facilitating the development of smart cyber-

physical systems.

2.3 XAI

Explainable AI (XAI) is crucial in addressing the lack

of transparency in AI-driven CPS recommendation

systems. XAI aims to make AI decisions comprehen-

sible to humans by providing clear insights into how

models generate predictions (Ribeiro et al., 2016b).

In the IoT-AID system, post-hoc interpretability

methods such as Local Interpretable Model-agnostic

Explanations (LIME) and SHapley Additive exPla-

IoT-AID: Leveraging XAI for Conversational Recommendations in Cyber-Physical Systems

673

nations (SHAP) have been incorporated to enhance

transparency (Ribeiro et al., 2016a; Garouani and

Bouneffa, 2023). LIME generates localized expla-

nations by approximating decision boundaries around

individual predictions, enabling users to understand

the key factors inﬂuencing each recommendation. On

the other hand, SHAP provides a global interpretabil-

ity framework by calculating the contribution of each

input feature across multiple predictions. These tech-

niques empower users to validate the CPS recommen-

dations, fostering greater trust and usability.

3 A NOVEL CYBER PHYSICAL

RECOMMENDER SYSTEM

3.1 System Architecture

IoT-AID is designed as a modular system compris-

ing several key components that work in tandem to

streamline CPS design and conﬁguration. The ﬁrst

component, the Preprocessing Module, is responsi-

ble for cleaning and tokenizing text input using nat-

ural language processing tools such as spaCy and

NLTK. This ensures that the input is standardized

and free from noise. Next, the NLP module uti-

lizes ﬁne-tuned BERT embeddings to extract relevant

entities and contextual information from the input.

The extracted entities are then processed in the data

analysis module, which leverages a combination of

traditional machine learning models, deep learning

architectures, and transformer-based models to ana-

lyze user requirements comprehensively. Based on

this analysis, the Recommendation Engine generates

ranked CPS conﬁgurations tailored to the user’s spe-

ciﬁc application needs. Finally, the Explainability

Layer integrates XAI techniques such as LIME and

SHAP to provide transparent insights into the recom-

mendations, allowing users to understand the reason-

ing behind each decision.

3.2 Dataset Generation and Feature

Extraction

Data Acquisition Challenges: In our pursuit of con-

structing a reliable Cyber-Physical Recommender

System (CPRS) as shown in ﬁgure 2, we encountered

a signiﬁcant challenge: the absence of readily avail-

able data online. To address this gap, we turned to

web scraping and generative AI techniques to gener-

ate the necessary data.

Dataset Composition: The dataset needed to in-

clude a variety of applications, their respective ﬁelds

of activity, and the sensors used in such applications.

To accomplish this, we integrated a function into our

Django project. This function leverages the OpenAI

library and an OpenAI API key to make requests to

the ChatGPT API, allowing us to retrieve the data we

required.

Data Generation Process: We initiated the data

generation by passing a speciﬁc phrase as a param-

eter in the request, instructing ChatGPT to provide us

with the needed data. After executing the request, we

successfully obtained several lines of data. However,

we faced a limitation: ChatGPT could only provide

limited lines per request, and we needed hundreds.

To overcome this, we ran the request in a loop 100

of times, storing all the data in JSON format within a

JSON ﬁle.

Data Redundancy Issue: This process took sev-

eral hours, and we eventually collected all required

JSON elements. Yet, we encountered an issue with

data redundancy—too many repeated lines. Despite

instructing ChatGPT not to provide previously given

applications, the data still contained duplicates.

Redundancy Resolution: To resolve this, we im-

plemented a function that iterated through the JSON

array, ﬁltering out redundant entries and compiling a

new ﬁle with unique elements. Once we had a reﬁned

JSON ﬁle containing 600 unique applications, we uti-

lized a CSV library to convert the JSON data into a

CSV format.

This structured approach not only streamlined our

data collection process but also ensured the unique-

ness and relevance of the data for our AI model’s

training.

3.2.1 Entities Dataset

The entities are described by the attributes shown in

ﬁgure 3 :

• Application: This ﬁeld speciﬁed the context or

domain where sensors were applied.

• Sensor Types: It listed the various types or cate-

gories of sensors commonly used within the spec-

iﬁed domain.

• Description: This section provided a detailed

overview of how sensors were employed within

the speciﬁed domain, including their functionali-

ties and practical applications.

• Domain : Here, we identiﬁed the broader industry

or ﬁeld where these sensors found utility.

3.3 Data Preprocessing

Data preprocessing is a critical step in preparing the

text data for machine learning models. We utilized

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

674

Data Collection

Data Genration

Data set

Data

PreProcessing

Feature Extraction

Traditional ML

Decision Tree

SVM

Naive Bayes

Deep Learning

RNN-LSTM

CNN

Fine-Tuning

Bert

GPT

Combining

Predictions

Prediction Explainability

Output

(prediction+Explainability)

Figure 2: CPRS Pipeline.

Figure 3: Entities Dataset.

both NLTK and spaCy for this purpose. The process

involved cleaning the text by removing special char-

acters, numbers, and stop words which do not con-

tribute to the semantic meaning. Tokenization was

performed to split the text into individual words (to-

kens). This was followed by lemmatization, which re-

duced words to their base forms, thus ensuring consis-

tency (Garouani and Kharroubi, 2022). For instance,

the raw text ”I need a system for health monitoring

for heart rate and oxygen levels” was preprocessed

using spaCy to become ”need system health monitor-

ing heart rate oxygen level”. This preprocessing en-

sured that the text was in a suitable format for feature

extraction and modeling.

3.4 Feature Extraction

Feature extraction transforms textual data into nu-

merical representations that can be used by machine

learning algorithms. We employed two primary meth-

ods: TF-IDF (Term Frequency-Inverse Document

Frequency) Vectorization and BERT Embeddings.

TF-IDF vectorization converted the text into numeri-

cal features by considering the importance of words in

the context of the entire dataset. BERT embeddings,

generated using a pre-trained BERT model, provided

deep contextual understanding of the text by captur-

ing the semantic relationships between words. For

example, the preprocessed text ”need system health

monitoring heart rate oxygen level” was converted

into TF-IDF vectors and BERT embeddings, which

served as inputs to the various models.

3.5 Model Training

3.5.1 Machine Learning Models

• Decision Tree classiﬁer (DT): The Decision Tree

classiﬁer is a simple yet powerful model that splits

the data based on feature importance to make pre-

dictions. We trained a single decision tree on TF-

IDF vectors using sklearn’s Decision Tree classi-

ﬁer. The model was capable of capturing com-

plex decision boundaries but was prone to over-

ﬁtting. Despite this, it provided a clear and inter-

pretable decision-making process, which we visu-

alized through a generated decision tree diagram.

• Support Vector Machine (SVM): The SVM model

with a linear kernel is well-suited for text classi-

ﬁcation due to its robustness and ability to gen-

eralize well. We implemented an SVM classiﬁer

within a pipeline that included TF-IDF vectoriza-

tion, using sklearn’s SVC with probability esti-

mates. This model transformed the preprocessed

descriptions into numerical vectors and then clas-

siﬁed them. The SVM showed high accuracy on

validation data and was less prone to overﬁtting

compared to the Decision Tree.

• Naive Bayes: The Naive Bayes classiﬁer, partic-

ularly the Multinomial Naive Bayes, is efﬁcient

and effective for text data. It makes simple as-

sumptions about the data distribution but performs

well. Using sklearn’s MultinomialNB, we trained

the model on TF-IDF vectors. The NB model was

fast to train and predict, and it achieved good per-

formance on the text classiﬁcation task.

3.5.2 Deep Learning Models

• Recurrent Neural Network (RNN): RNNs are de-

signed to handle sequential data and are particu-

larly effective for tasks involving time series or

IoT-AID: Leveraging XAI for Conversational Recommendations in Cyber-Physical Systems

675

text sequences. We built an RNN using a sequen-

tial model with an Embedding layer, an LSTM

layer, and a Dense output layer using TensorFlow.

The model captured sequential dependencies in

the text, allowing it to understand context over

long sequences. We tokenized and padded the in-

put sequences before feeding them into the RNN.

This model achieved good accuracy on the valida-

tion data, demonstrating its ability to handle com-

plex dependencies in the text.

• Convolutional Neural Network (CNN): CNNs, al-

though traditionally used for image processing,

are also effective for text classiﬁcation by cap-

turing local patterns. Our CNN model consisted

of an Embedding layer, a Conv1D (convolutional)

layer, a GlobalMaxPooling1D layer, and a Dense

output layer. This architecture allowed the model

to detect key phrases and patterns within the text.

We tokenized and padded the sequences before

feeding them into the CNN. The model showed

high accuracy on the validation data, effectively

capturing local features in the text.

• Fine-Tuning BERT:Fine-tuning a pre-trained

BERT model involves adapting it to our speciﬁc

text classiﬁcation task. BERT is a transformer-

based model that excels at understanding context

in text. For our task, we used the pre-trained

BERT model from the Hugging Face library and

added a classiﬁcation layer on top of it. This

involved training the entire model, including the

BERT layers, to adjust the pre-trained weights

to better ﬁt our dataset. We tokenized the input

text using BERT’s tokenizer, ensuring compatibil-

ity with the model’s expectations. The tokenized

inputs were then fed into the BERT model, fol-

lowed by a dense layer for classiﬁcation. Train-

ing was performed using a learning rate schedule

and early stopping to prevent overﬁtting. The ﬁne-

tuning process leveraged the deep contextual em-

beddings of BERT, making it highly effective for

our classiﬁcation task.

3.6 Proposed Model for XAI in CPS

3.6.1 Role of XAI

Explainable AI (XAI) is fundamental to the IoT-AID

system, ensuring that its recommendations are not

only accurate but also comprehensible. The impor-

tance of XAI lies in its ability to make AI systems

transparent by explaining the reasoning behind their

outputs. For example, in a CPS design scenario,

XAI can clarify why speciﬁc sensors or communi-

cation protocols were chosen for a given applica-

tion (Garouani and Bouneffa, 2023). Such explana-

tions enable engineers to validate the system’s rec-

ommendations and align them with real-world con-

straints and requirements. Moreover, by building trust

and fostering informed decision-making, XAI miti-

gates skepticism often associated with AI-driven sys-

tems.

3.6.2 Utilization of XAI

IoT-AID incorporates XAI techniques at various

stages of its workﬂow. During preprocessing, Local

Interpretable Model-agnostic Explanations (LIME)

and SHapley Additive exPlanations (SHAP) help

identify the most signiﬁcant features in user inputs.

LIME generates local approximations of the model’s

decision boundaries, allowing users to understand

which factors contributed most to speciﬁc recommen-

dations. SHAP, on the other hand, provides a global

explanation by computing the average contribution of

each input feature across multiple predictions. These

techniques provide localized insights, making the sys-

tem’s behavior transparent for speciﬁc recommenda-

tions. Additionally, during entity recognition and rec-

ommendation generation, XAI techniques highlight

the role of extracted features and contextual rela-

tionships in shaping system outputs. This layered

approach to interpretability ensures that every stage

of IoT-AID’s process is comprehensible and user-

centric.

3.6.3 Structure of the Model

The IoT-AID system integrates traditional machine

learning and modern deep learning techniques, com-

plemented by XAI tools. Traditional algorithms,

such as Decision Trees and Support Vector Machines

(SVMs), are employed for their inherent interpretabil-

ity and simplicity. For more complex tasks, Con-

volutional Neural Networks (CNNs) and Recurrent

Neural Networks (RNNs) are used to capture intricate

relationships and temporal dependencies in the data.

Fine-tuned transformer-based models, such as BERT,

enable contextual understanding and domain-speciﬁc

entity recognition. These models utilize structured

CPS datasets and expert-generated synthetic data to

improve recommendation accuracy and robustness.

The architecture ensures that each layer processes

data efﬁciently, applying XAI techniques to provide

transparent insights at every stage.

3.6.4 Model Architecture and Workﬂow

IoT-AID’s architecture consists of ﬁve key modules:

Preprocessing Module: This module cleans and

tokenizes user inputs, removing noise and standardiz-

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

676

ing the data format. Tools like spaCy and NLTK are

employed for efﬁcient text preprocessing.

NLP Module: Using ﬁne-tuned BERT embed-

dings, this module extracts entities and semantic con-

text from user inputs, ensuring accurate interpretation

of requirements.

Data Analysis Module: Outputs from various ma-

chine learning and deep learning models are inte-

grated in this module to evaluate user requirements

comprehensively.

Recommendation Engine: Based on the processed

data, this engine generates ranked CPS conﬁgurations

tailored to the user’s application.

Explainability Layer: XAI techniques are ap-

plied here to provide transparent explanations for rec-

ommendations, fostering user trust and enabling in-

formed decision-making. Figure 4 shows a snippet of

how we used Lime highlighting key words that inﬂu-

ence the model’s decision, with important terms in red

and less important terms in green.

Figure 4: Lime Explanation.

To give an example of the work ﬂow, the user be-

gins by providing a query, such as ”I need a system

for health monitoring for heart rate and oxygen lev-

els.” This input is cleaned and tokenized in the Pre-

processing Module before being passed to the NLP

Module, where relevant entities like ”health monitor-

ing” and ”heart rate” are extracted. These entities

are analyzed in the Data Analysis Module using a

combination of machine learning and deep learning

models, which collectively generate a list of recom-

mended conﬁgurations. The Recommendation En-

gine ranks these suggestions, while the Explainabil-

ity Layer provides detailed insights into the reasoning

behind each recommendation. The models use struc-

tured CPS datasets combined with expert-generated

synthetic data to improve recommendation accuracy

and robustness. The architecture also supports iter-

ative feedback loops, allowing users to reﬁne their

requirements and improve the model’s recommenda-

tions over time.

4 RESULTS AND DISCUSSION

Preliminary evaluations of IoT-AID have demon-

strated its effectiveness in addressing key challenges

in CPS design. The system achieved high accuracy

in entity recognition and recommendation relevance,

thanks to its integration of advanced language models

and machine learning techniques. The incorporation

of XAI methods signiﬁcantly enhanced user trust and

satisfaction, providing clear and concise explanations

for recommendations. However, the system’s perfor-

mance remains constrained by the quality and scope

of the training dataset, highlighting the need for fur-

ther data expansion.

Also some challenges with data scarcity arise:

4.1 Data Collection Issues

CPS applications require domain-speciﬁc datasets,

which are often unavailable or incomplete. IoT-AID

addresses this challenge by leveraging web scrap-

ing techniques to collect data from publicly available

sources. Additionally, generative AI models such as

OpenAI’s GPT are used to synthesize supplementary

data, bridging gaps in the dataset.

4.2 Data Preprocessing and

Deduplication

The data collection process is often plagued by redun-

dancy, with duplicate entries diminishing the qual-

ity of the dataset. To counter this, IoT-AID employs

automated scripts for deduplication and enriches the

dataset with relevant attributes, such as sensor types

and application domains. These steps ensure that the

data is both comprehensive and unique.

4.3 Impact on Results

While these methods signiﬁcantly enhance dataset

quality, data scarcity remains a challenge in niche

CPS domains. The system’s performance is partic-

ularly affected in scenarios requiring highly special-

ized knowledge. Future work will explore collabora-

tions with industry partners and the use of advanced

data augmentation techniques to further expand and

diversify the dataset.

4.4 Evaluating CPS Performance and

Decision-Making with XAI

Evaluating CPS performance requires a problem-

speciﬁc approach, as different applications necessi-

tate distinct success metrics. A universal evaluation

criterion is impractical, so future work will focus on

deﬁning tailored performance indicators or relying on

domain expert validation. This approach ensures that

reasoning strategies—whether pattern-based, logic-

driven, or hybrid—are aligned with the intended func-

tion of the CPS.

IoT-AID: Leveraging XAI for Conversational Recommendations in Cyber-Physical Systems

677

In addition, it is important to recognize that

explainability in AI-driven decision-making is not

equivalent to correctness. XAI in IoT-AID is de-

signed to enhance user understanding by providing

interpretable recommendations rather than guarantee-

ing optimal solutions. By offering clear justiﬁcations

for each recommendation, users can make more in-

formed decisions, either accepting or rejecting sug-

gestions based on their own expertise and contextual

needs.

4.5 Addressing Bias Reinforcement in

Recommendations

A signiﬁcant challenge in AI-driven recommenda-

tions is the risk of bias reinforcement. If an XAI sys-

tem prioritizes user trust and alignment over objec-

tive accuracy, it may reinforce predictable but subop-

timal decisions. Future work will focus on develop-

ing mechanisms to detect and mitigate biases, such as

adversarial testing, diverse training datasets, and al-

ternative recommendation strategies. These measures

aim to ensure that recommendations remain explain-

able while also being objectively beneﬁcial for CPS

conﬁgurations.

5 CONCLUSION AND FUTURE

WORK

IoT-AID represents a signiﬁcant advancement in the

ﬁeld of CPS design and implementation. By inte-

grating XAI techniques into a comprehensive recom-

mendation system, IoT-AID addresses critical chal-

lenges related to complexity, data scarcity, and trans-

parency. Its ability to provide interpretable, accurate,

and user-centric recommendations has the potential to

democratize CPS adoption and accelerate the realiza-

tion of Industry 4.0 objectives. Future iterations will

reﬁne the system’s capabilities, ensuring its applica-

bility across diverse industries and domains.

Future efforts will focus on several key areas to

enhance IoT-AID’s capabilities. First, expanding the

dataset by collaborating with industry partners and

employing advanced data augmentation techniques

will improve model accuracy and generalizability.

Second, the exploration of hybrid XAI techniques that

combine intrinsic and post-hoc interpretability meth-

ods will further enhance transparency. Third, ﬁne-

tuning advanced transformer models, such as GPT

variants, for domain-speciﬁc applications will enable

more nuanced and accurate recommendations. Fi-

nally, real-world deployments of IoT-AID in indus-

trial settings will provide valuable insights into its

scalability, adaptability, and overall impact.

REFERENCES

Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D.,

Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G.,

Askell, A., et al. (2020). Language models are few-

shot learners. Advances in Neural Information Pro-

cessing Systems, 33:1877–1901.

Choaib, M., Garouani, M., Bouneffa, M., Waldhoff, N.,

and Mohanna, Y. (2024). Iot-aid: An automated deci-

sion support framework for iot. SN Computer Science,

5(4):429.

Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K.

(2019). Bert: Pre-training of deep bidirectional trans-

formers for language understanding. In Proceedings

of the 2019 conference of the North American chapter

of the association for computational linguistics: hu-

man language technologies, volume 1 (long and short

papers), pages 4171–4186.

Garouani, M., Ahmad, A., Bouneffa, M., and Hamlich, M.

(2022a). Amlbid: An auto-explained automated ma-

chine learning tool for big industrial data. SoftwareX,

17:100919.

Garouani, M., Ahmad, A., Bouneffa, M., and Hamlich,

M. (2023). Autoencoder-knn meta-model based data

characterization approach for an automated selection

of ai algorithms. Journal of Big Data, 10(1).

Garouani, M., Ahmad, A., Bouneffa, M., Hamlich, M.,

Bourguin, G., and Lewandowski, A. (2022b). Towards

big industrial data mining through explainable auto-

mated machine learning. The International Journal of

Advanced Manufacturing Technology, 120(1):1169–

1188.

Garouani, M. and Bouneffa, M. (2023). Unlocking the

black box: Towards interactive explainable automated

machine learning. In International Conference on In-

telligent Data Engineering and Automated Learning,

pages 458–469. Springer.

Garouani, M., Hamlich, M., Ahmad, A., Bouneffa, M.,

Bourguin, G., and Lewandowski, A. (2022c). To-

ward an Automatic Assistance Framework for the Se-

lection and Conﬁguration of Machine Learning Based

Data Analytics Solutions in Industry 4.0, page 3–15.

Springer International Publishing.

Garouani, M. and Kharroubi, J. (2022). Towards a New

Lexicon-Based Features Vector for Sentiment Anal-

ysis: Application to Moroccan Arabic Tweets, page

67–76. Springer International Publishing.

Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep

learning. MIT press.

Hochreiter, S. and Schmidhuber, J. (1997). Long short-term

memory. Neural computation, 9(8):1735–1780.

Joachims, T. (1998). Text categorization with support vec-

tor machines: Learning with many relevant features.

European Conference on Machine Learning, pages

137–142.

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

678

Mardiansyah, K. and Surya, W. (2024). Comparative analy-

sis of chatgpt-4 and google gemini for spam detection

on the spamassassin public mail corpus.

McCallum, A. and Nigam, K. (1998). A comparison of

event models for naive bayes text classiﬁcation. AAAI

workshop on learning for text categorization, 752:41–

48.

Moosavi, S., Farajzadeh-Zanjani, M., Razavi-Far, R.,

Palade, V., and Saif, M. (2024). Explainable ai in

manufacturing and industrial cyber–physical systems:

a survey. Electronics, 13(17):3497.

Quinlan, J. R. (1986). Induction of decision trees. Machine

learning, 1(1):81–106.

Ribeiro, M. T., Singh, S., and Guestrin, C. (2016a). Model-

agnostic interpretability of machine learning. arXiv

preprint arXiv:1606.05386.

Ribeiro, M. T., Singh, S., and Guestrin, C. (2016b). Why

should i trust you?: Explaining the predictions of any

classiﬁer. Proceedings of the 22nd ACM SIGKDD In-

ternational Conference on Knowledge Discovery and

Data Mining, pages 1135–1144.

Sebastiani, F. (2002). Machine learning in automated

text categorization. ACM computing surveys (CSUR),

34(1):1–47.

Whang, S. E., Roh, Y., Song, H., and Lee, J.-G. (2023).

Data collection and quality challenges in deep learn-

ing: A data-centric ai perspective. The VLDB Journal,

32(4):791–813.

Yang, Z., Dai, Z., Yang, Y., Carbonell, J., Salakhutdinov,

R. R., and Le, Q. V. (2019). Xlnet: Generalized au-

toregressive pretraining for language understanding.

Advances in neural information processing systems,

32.

Zaheer, M., Guruganesh, G., Dubey, K. A., Ainslie, J., Al-

berti, C., Ontanon, S., Pham, P., Ravula, A., Wang,

Q., Yang, L., et al. (2020). Big bird: Transformers

for longer sequences. Advances in neural information

processing systems, 33:17283–17297.

Zampieri, M., Rosenthal, S., Nakov, P., Dmonte, A., and

Ranasinghe, T. (2023). Offenseval 2023: Offensive

language identiﬁcation in the age of large language

models. Natural Language Engineering, 29(6):1416–

1435.

Zhao, Z. and Wu, Y. (2016). Attention-based convolutional

neural networks for sentence classiﬁcation. In Inter-

speech, volume 8, pages 705–709.

IoT-AID: Leveraging XAI for Conversational Recommendations in Cyber-Physical Systems

679