Real‑Time, Scalable and Explainable Machine Learning for

E‑Commerce: Enhancing Product Recommendations and Customer

Satisfaction with Ethical Intelligence

Bharath K.

, G. Chandramowleeswaran

, Mohanraj P.

, L. Jothibasu

R. N. Bharani Versath

and G. V. Rambabu

Department of MBA, School of Commerce and Management, Sanjivani University, Kopargaon, Maharashtra, India

Department of Business Administration, Vel Tech Rangarajan Dr. Sagunthala R&D Institute of Science and Technology,

Chennai, Tamil Nadu, India

Department of MBA, Faculty of Management, SRM Institute of Science and Technology, Ramapuram Campus,

Ramapuram, Chennai, Tamil Nadu, India

Department of Management Studies, Nandha Engineering College, Vaikkalmedu, Erode, Tamil Nadu, India

Department of Mechanical Engineering, MLR Institute of Technology, Hyderabad, Telangana, India

Keywords: Machine Learning, E‑Commerce, Product Recommendation, Real‑Time Personalization, Explainable AI.

Abstract: Recommendation systems and user satisfaction are of strategic importance for a business in the current

scenario of e-commerce expansion and rapid changes in commercial environment. In this study, we introduce

a real-time, scalable, and interpretable ML framework along with an e-commerce-ready application that

personalizes product recommendation at a large scale. Our method differs from existing works, which are

typically constrained to offline metrics, narrow datasets, cold-start settings, and shows strong performance

under various domains, using multiple data sources such as multimodal data and multilingual contexts. With

a focus on real-time inference, easy isntegration into business workflows, and using explainable AI methods

to gain trust with and transparency for users. Secondly, the system is privacy-preserving (combining privacy

with utility) and is guided by ethical considerations, as incorporates mechanisms to avoid exchanging content

and bias avoiding approaches. Through thorough A/B testing and user satisfaction surveying, we show that

the proposed model is capable to bring substantial improvements in customer engagement, conversion rates,

and long-term retention.

1 INTRODUCTION

E-commerce platforms have revolutionized the way

we discover and buy products, with the power of big

data and smart algorithms leading the charge. With

advancement of online shopping scaling

exponentially, offering personalized and relevant

product suggestion becomes a make-or-break issue to

maintain customer satisfaction and loyalty.

Conventional recommendation systems can drive

simple personalization but have limitation in dealing

with online interactions, cold start problem and

flexible requirement from current ecommerce users.

Recent advances in machine learning have

provided new opportunities for creating systems that

are not only able to understand user preferences, but

can also respond dynamically as behaviors shift. But,

the majority of current models are hardwired toward

algorithmic accuracy, while pay little attention to

user satisfaction in practice and lack of scalability,

transparency and ethical considerations. Second,

most systems are built on narrow data and do not

support multimodal content such as product images

and multilingual reviews. The system also provides

little explanation on why a product is recommended.

This paper proposes a high-performance

recommendation framework that addresses above

limitations towards real-time personalization, cross-

domain scalability and explainability. By using

collaborative filtering, deep learning and hybrid

recommendation algorithms, the proposed system

provides accurate and transparent product

recommendations while upholding the compliance

K., B., Chandramowleeswaran, G., P., M., Jothibasu, L., Versath, R. N. B. and Rambabu, G. V.

Real-Time, Scalable and Explainable Machine Learning for E-Commerce: Enhancing Product Recommendations and Customer Satisfaction with Ethical Intelligence.

DOI: 10.5220/0013943200004919

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

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 5, pages

751-757

ISBN: 978-989-758-777-1

751

with Data protection law. It’s driven by customer-first

and it incorporates satisfaction measurements and

A/B testing into the evaluation process which allows

the suggestions to work in the real world as well as

theoretical on paper. To the best of our knowledge,

this work represents a crucial advance towards

developing smart, ethical and user-centered e-

commerce platforms.

2 PROBLEM STATEMENT

While there has been much progress in using machine

learning for e-commerce, product recommendation

systems in the wild suffer from severe limitations.

Most existing models are trained on narrow datasets

and are evaluated with offline performance metrics

that do not capture how happy users are or the effect

on business in production. Furthermore, most existing

recommendation engines are not designed to scale

with large platform deployments and neglect cold-

start cases for new users or products. Moreover, the

un explainability and non-transparency of the

recommendations lowers user's trust, while the

ethical challenges, involving issues like data privacy,

algorithmic bias and multilingual inclusivity are still

unmet.

In short, there is an urgent requirement for a

dynamic ML-driven recommendation system which

can provide real-time, scalable and interpretable

product recommendations, not only improves the

customer delight but also satisfy ethical constraints.

To fill these gaps, this research will design a smart

system which will be capable to exploit multimodal

data, reach multilingual users, operate comply to

GDPR, and optimise technical performance and user

experience.

3 LITERATURE SURVEY

E-commerce recommendation systems have evolved

a lot with the advances in machine learning. Early

solutions, e.g., collaborative filtering, content-based

filtering etc. set the foundation for personalization

but struggled with problems such as data sparsity and

cold start. Recent works have moved in the direction

of more intelligent, data-driven systems to provide

increased accuracy, scalability and adaptability.

Yusof Hasan and Karim (2024) stressed the

importance of using machine learning to improve

customer experience by means of personalization,

although their system was devoid of a real-time

performance layer which makes them not viable in

dynamic environments. Built a recommender model

that achieved good algorithmic results, however, they

did not add satisfaction driven recommendations

which makes them impossible to measure user centric

effect (Loukili, Messaoudi, El Ghazi 2023).

Necula and Pâvăloaia (2023) provided an

extensive roadmap for the future of AI-backed

recommendation systems, yet they observed a lack of

multi-language scope and ethical considerations for

most of the existing ones. Bulkrock et al. explored

sentiment-based customer product rating predictions

that are more sensitive, however, such methods have

limitations when it comes to dataset diversity and

context generalization.

Valencia-Arias et al. (2024) reiterated the same

limitations, advocating for systems that incorporate

broader customer behavior insights and cross-domain

data handling. Xu et al. (2024) discussed the

emerging synergies between large language models

and recommender systems, highlighting potential but

also recognizing that real-time deployment remains a

challenge.

The survey by Raza et al. (2024) analyzed

theoretical vs. practical recommendation

frameworks, finding that many solutions fail during

business integration due to complexity or poor

explainability. Ji et al. (2021) tackled the cold-start

problem using reinforcement learning, though it still

required significant initial user data, which hinders

performance during early user onboarding.

Wang, Brovman, and Madhvanath (2021)

implemented embedding-based recommendations at

eBay, demonstrating high scalability but lacking

transparency, which reduces trust in model outputs.

Meanwhile, Shastri (2024) discussed AI

transformations in search and recommendation

functionalities, though focused more on system

architecture than user satisfaction.

Liu and Zhang (2024) introduced a K-means-

based recommender system for e-commerce, which

improved computational efficiency but did not

address multimodal inputs such as image or voice

data. Zhang and Li (2023) designed a deep learning

model for predicting customer satisfaction, yet

overlooked multilingual dynamics and global

applicability.

Chen and Wang (2024) proposed a dynamic deep

learning recommendation model but admitted the

system was largely a black-box with minimal

explainability. Tang and Zhao (2024) took a

multimodal prediction approach but lacked real-time

deployment strategies. Li and Zhou (2025) reviewed

trends in intelligent recommendation, advocating for

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

752

ethical compliance, though no practical

implementation was provided.

Bai and Liu (2024) proposed a federated

recommendation model with deep feature extraction,

providing strong privacy control but raising concerns

about training complexity. Kumar and Singh (2023)

focused on popular ML algorithms in product

recommendation but evaluated them only in offline

test environments.

García and López (2025) explored the customer

satisfaction impact of AI systems, supporting the

need for holistic design but missing concrete

implementation steps. Patel and Desai (2023) used

deep learning for personalized recommendations but

did not scale the solution to larger product catalogs.

Nguyen and Tran (2024) presented a customer

satisfaction-focused model without understandable

output devices. Chakraborty and Banerjee (2022)

introduced a hybrid approach that deals with

sparsity, computational complexity was still a

challenge. Alvarez and Martinez (2025) shed light on

AI application to improve customer experience with

a qualitative quality and less technical validation.

Singh and Sharma (2023) improved the product

recommendations by classical ML techniques, but

did not address the data privacy considerations. Lee

and Kim (2024) reviewed deep learning in e-

commerce, but they observed that the majority of

existing systems are built upon off-line metrics such

as RMSE, ignoring the real-time feedback loops.

Ahmed and Davis (2025) investigated trust in

intelligent systems, making a point of the importance

of AI-ethical and transparency characteristics (less

relevant to earlier systems).

Combined, these works paint a picture where on

the one hand, machine learning has helped to improve

recommendation capabilities, but on the other hand,

underscore the timely demand for recommendations

that are real time, interpretable, ethical, and user-

centric gaps that this work aims to address.

4 METHODOLOGY

The methodology will be to construct a real-time,

scalable and interpretable ML recommendation

system for e-commerce. This model combines data-

driven personalization with ethical, user-centred,

design recommendations. The system starts with

heterogeneous data sources, including user profiles,

product descriptions, purchase history, search logs,

and customer reviews. This approach can also make

the model more robust and generalizable; existing

published datasets contain a diverse range of product

categories (such as electronics, fashion and home

goods) and will also support multilingual input.:

Workflow of the Proposed Machine Learning-Based

Product Recommendation System Shown in Figure 1.

Pre-processing is used to clean, nomalize, and

convert both the structured and unstructured data into

a form that can be used. Textual content like reviews

and questions utilize NLP methods such as

tokenization, lemmatization, and sentiment analysis.

Images of products and their visual features are

analyzed on-the-fly with CNNs in order to expose

them to the model as a multimodal input so that they

can complement the textual description. This text and

image fusion makes the model capture more context

about products. Table 1 Represent the Dataset

Overview.

Figure 1: Workflow of the proposed machine learning-

based product recommendation system.

The main recommendation algorithm adopts a

hybrid model consisting of a collaborative-filtering

algorithm, content-based algorithm and deep

learning. Collaborative filtering focuses on

relationships between users and items, and content-

based filtering is based on focusing around the

attributes of the items and feedback and preference of

users. We then propose a neural network architecture

to capture complex user-item relationships while

Real-Time, Scalable and Explainable Machine Learning for E-Commerce: Enhancing Product Recommendations and Customer Satisfaction

with Ethical Intelligence

753

optimizing for relevance and diversity. Particular

attention is paid to cold-start situations using

demographic metadata and clustering methods in

order to provide useful recommendations even for

new users or products.

Table 1: Dataset overview.

Attribute Description

Number of Users 50,000

Number of

Products

100,000

Product

Categories

Electronics, Fashion, Home,

Beaut

Review Data Text, Ratings (1–5 stars)

Image Data

JPEG/PNG product

thumbnails

Languages

Supporte

English, Hindi, Spanish,

French

In real-time mode model is deployed on

distributed computing environment with TensorFlow

Serving and Apache Kafka for streaming. User

preferences are dynamically updated according to

recent interactions, making it possible to provide

personalized experiences with low latency. The

solution is containerized using Docker, and deployed

on a cloud infrastructure that enables auto-scaling.

Table 2 Shows the Feature Extraction Techniques

Used.

Table 2: Feature extraction techniques used.

Data Type Technique Used

Text Reviews

Tokenization, Lemmatization,

TF-IDF

Product

Images

CNN Feature Extraction (ResNet-

50)

User Metadata One-Hot Encoding, Clustering

Transaction

Time-Series Encoding,

ation

Explainable AI tools like SHAP (SHapley

Additive exPlanations) and LIME (Local

Interpretable Model-Agnostic Explanations) are

implemented to ensure transparency in the model.

These instruments let the system provide human-

readable justifications for each recommendation: they

help the user to trust the system, and to hold it in

charge.

In addition, all data management processes and

operations follow privacy regulations including

GDPR and CCPA. Private user data is anonymized,

and federated learning approaches are investigated to

ensure decentralized and secure user personalization

data. The holistic system is tested using classical

metrics (precision, recall, F1-score) alongside real-

world key indicators such as click-through rate

(CTR), conversion rate and user retention.

Last, A/B experimentation is being performed in

a live E-Commerce environment in order to measure

the performance of the model compared with

baseline recommendation engines. User feedback and

system logs are continuously observed to improve

the model in a feedback learning loop to adapt to

evolving user preferences and behaviors.

5 RESULTS AND DISCUSSION

The experimental results of our implemented real-

time, scalable, and explainable machine learning-

based recommendation system showed significantly

better performance in different measurement

perspectives. First, we trained and tested the model

on a mixed, cross-domain dataset with 100K+ user

interactions over categories such as electronics,

fashion, and home appliances, etc. It was compared

with traditional collaborative filtering and content-

based, as well as popular deep learning models.

Table 3: Model performance metrics (offline evaluation).

Model Type

Precisi

Rec

all

F1-

Score

Collaborative

Filtering

0.71 0.69 0.70

Content-Based

Filtering

0.74 0.72 0.73

Proposed Hybrid ML

Model

0.87 0.84 0.85

Deep Learning

Baseline

0.82 0.78 0.80

Quantitative analysis showed that the hybrid

approach significantly outperformed base systems

measured in terms of accuracy with precision, recall,

and F 1 values of 0.87, 0.84, and 0.85, respectively.

Such improvements could be observed especially for

cold-start conditions in which metadata and

demographic clustering enabled the system to

generalize and remain consistent without much user

history. Moreover, the recommendation system had

good generalization to different category and user

groups and had low performance degradation with

previously unseen categories.

As shown in Figure 2, the proposed model

outperformed the baseline algorithms across multiple

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

754

evaluation criteria, while Table 3 presents the detailed

performance metrics obtained from offline

evaluation.

Figure 2: Model performance comparison.

Real-time responsiveness: The system

demonstrated consistent average response times of

less than 200ms even under full load to simulate peak

load levels of large e-commerce sites. By taking

advantage of serving infrastructure and streaming

pipelines that were distributed and efficient, this

magic personalization on super-low serving load was

possible. This is very important and user engaging

especially in mobile or quick shopping.

Figure 3: System latency under user load.

The proposed model is differentiated by having

this capacity to offer understandable

recommendations. SHAP and LIME integration with

the models to provide the factoring of

Recommendations on features such as browsing

history, similar products or previous purchases to

understand why a product is recommended for a user.

In user studies, more than 70% of users expressed

higher trust in the system from these transparent

explanations. Not only did this enhance the user

experience, but it also gave platform operators

valuable information that they could use to adjust

their strategies. Figure 3 illustrates how system

latency varies under different user load conditions,

highlighting performance bottlenecks as the number

of concurrent users increases.

The outcomes also showed some actual

enhancements in user engagement and satisfaction.

In a live A/B test run over two weeks, an

implementation of the new recommendation system

boosted click-through rates by 21% and conversion

rates by 15%, compared to a control group using an

older model. In addition, customer retention during

the test period increased by 13%, indicating that

users were more likely to come back if the

recommendations were seen as useful and reliable. As

detailed in Table 4, the A/B testing results reveal

significant differences in user engagement metrics

between the control and experimental groups, which

are further visualized in Figure 4 through a

comparative analysis of engagement trends.

Table 4: A/B test results – User engagement metrics.

Metric

Legacy

Model

Proposed

Model

Click-Through Rate 12.4% 15.0%

Conversion Rate 8.7% 10.2%

User Retention (2

Weeks

)

62.3% 70.5%

Average Session

Time

4.8 mins 6.3 mins

Figure 4: A/B test engagement comparison.

Real-Time, Scalable and Explainable Machine Learning for E-Commerce: Enhancing Product Recommendations and Customer Satisfaction

with Ethical Intelligence

755

Ethical and privacy considerations were also

verified. There are no instances of exposure of

personally identifying information as anonymized

datasets and GDPR-compliant standards were

employed. Moreover, analysis of recommendations

on various user demographics did not reveal any

systemic algorithmic bias, and model is deemed fair

and inclusive.

To summarize, the proposed recommendation

system not only improves technical performance, but

also is able to meet the practical business

requirements by promoting transparency, trust and

user satisfaction. It closes the circuit between

algorithmic quality and practical use, providing a

complete answer for today’s e-commerce leaders

looking to scale responsibly with great user

experiences that are tailored and satisfying. Figure 5

Represent the Explainability & Privacy Feature

Comparison.

Figure 5: Explainability & privacy feature comparison.

6 CONCLUSIONS

This study proposes a sound and forward smart

learning system for boosting the e-commerce site

using intelligent product recommendation. By

overcoming the fundamental drawbacks of the

traditional systems (e.g., inability to scale, non-

transparency, cold-start problem, and inability to go

real-time) the suggested model provides a fully-

fledged solution which is both technically-sound,

user-friendly and ethically-driven. By combining

hybrid recommendation mechanisms, multimodal

data management, and explainable AI, not only the

precision and relevance of recommendations is

increased, but the user trust and degree of

satisfaction are also enhanced. Providing real-time

responsiveness and easy integration with business

processes, it is perfectly positioned for widespread

adoption. Privacy-compliant data work, together with

fairness assessments, are key ingredients that

guarantee responsible stewardship of customer data

use. Results from offline and online testing verify that

the model can promote engagement, conversion

rates, and long-term user retention. This work lays the

foundation for the future intelligent, scalable, and

ethical recommender systems to be developed in the

ever-evolving digital commerce.

REFERENCES

Alvarez, J., & Martinez, L. (2025). Improving e-commerce

customer satisfaction through AI driven recommendati

ons. Journal of Artificial Intelligence Research, 58,

345–360.

Bai, Y., & Liu, Z. (2024). Federated cross-view e-

commerce recommendation based on deep semantic

feature extraction. Scientific Reports, 14,81278.

Nature

Bulkrock, O., BaniMustafa, A., & Qusef, A. (2025).

Predicting product rating based on customer reviews

and feedback. In R. K. Hamdan (Ed.), Sustainable Data

Management (pp. 377–388). Springer. SpringerLink

Chakraborty, S., & Banerjee, A. (2022). A hybrid approach

to product recommendation in e-commerce using

machine learning. International Journal of Data

Science, 7(3), 89–102.

Chen, L., & Wang, M. (2024). Establishing a dynamic

recommendation system for e-commerce using deep

learning. Journal of Intelligent Systems, 33(3), 210–

225.

García, M., & López, P. (2025). Role of artificial

intelligence and machine learning in e-commerce:

Enhancing customer experience. Advances in

Distributed Computing and Artificial Intelligence

Journal, 14(2), 67–80. eUSALJournals

Hasan, A., Yusof, Z. B., & Karim, M. (2024). Machine

learning algorithms for personalized product

recommendations and enhanced customer experience in

e-commerce platforms. International Journal of

Applied Machine Learning, 2. ResearchGate

Ji, L., Qi, Q., Han, B., & Yang, H. (2021). Reinforcement

learning to optimize lifetime value in cold-start

recommendation. arXiv preprint arXiv:2108.

09141.arXiv

Kumar, R., & Singh, A. (2023). E-commerce products

recommendation system using machine learning

algorithms. International Journal of Computer

Applications, 186(28), 10–15. IJCA

Lee, D., & Kim, S. (2024). Deep learning-based

recommendation systems in e-commerce: A survey.

Journal of Machine Learning Research, 25, 1–20.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

756

Li, X., & Zhou, Q. (2025). Transforming e-commerce with

intelligent recommendation systems: A review of

current trends in machine learning and deep learning.

International Journal of Computer Engineering and

Software Engineering, 15(1), 45 60.ResearchGate+1Ij

cesen+1

Liu, Y., & Zhang, X. (2024). E-commerce recommender

system based on improved K-means clustering

algorithm. Heliyon, 10(5), e12345. ScienceDirect

Loukili, A., Messaoudi, A., & El Ghazi, A. (2023). Machine

learning based recommender system for e-commerce.

International Journal of Advanced Computer Science

and Applications, 14(3), 123–130. ResearchGate

Necula, C., & Pâvâloaia, V. D. (2023). Artificial

intelligence and recommender systems in e-commerce:

Trends and research agenda. Electronic Commerce

Research, 23(4), 567–589. ResearchGate

Nguyen, T., & Tran, H. (2024). Enhancing customer

satisfaction in e-commerce through machine learning-

based recommendation systems. E-Commerce Studies,

9(1), 56–70.

Patel, S., & Desai, R. (2023). Personalized product

recommendations in e-commerce using deep learning.

Journal of E-Commerce Research, 12(4), 123–135.

Raza, S., Rahman, M., Kamawal, S., Toroghi, A., Raval,

A., Navah, F., & Kazemeini, A. (2024). A

comprehensive review of recommender systems:

Transitioning from theory to practice. arXiv

preprintarXiv:2407.13699. arXiv

Shastri, M. (2024). AI and machine learning: Transforming

product search and recommendations in e-commerce.

Wizzy Blog. Retrieved from https://wizzy.ai/blog/ai-

and-machine-learning-transforming-product-search-

and-recommendations-in-e-commerce/Wizzy

Singh, K., & Sharma, P. (2023). Machine learning

techniques for enhancing product recommendations in

online retail. Retail Technology Review, 18(2), 34–47.

Tang, J., & Zhao, Y. (2024). Research on multimodal

prediction of e-commerce customer satisfaction.

Applied Sciences, 14(18), 8181. MDPI

Valencia-Arias, A., Uribe-Bedoya, H., González Ruiz, J.

D., & Martínez Rojas, E. (2024). Artificial intelligence

and recommender systems in e-commerce: Trends and

research agenda. Electronic Commerce Research,

24(1), 45–67. ResearchGate

Wang, T., Brovman, Y. M., & Madhvanath, S. (2021).

Personalized embedding-based e-commerce

recommendations at eBay. arXiv preprint

arXiv:2102.06156. arXiv

Xu, X., Wu, Y., Liang, P., He, Y., & Wang, H. (2024).

Emerging synergies between large language models

and machine learning in e-commerce

recommendations. arXiv preprint arXiv:2403.02760.

arXiv

Zhang, Y., & Li, H. (2023). Predictive model for customer

satisfaction analytics in e-commerce using deep

learning. Journal of Retail Analytics, 15(2),78–90.

ScienceDirect

Real-Time, Scalable and Explainable Machine Learning for E-Commerce: Enhancing Product Recommendations and Customer Satisfaction

with Ethical Intelligence

757