Volume 10, Issue 12, December – 2025 International Journal of Innovative Science and Research Technology

ISSN No: -2456-2165 https://doi.org/10.38124/ijisrt/25dec1676

Machine Learning for Personalized Fashion

Recommendation Systems: A Review
Harisu Aliyu1; Abdulamalik Abdulsalam2; Jamilu Musa3;
Grace Ojochenemi Emmanuelanorue4; Umar Muhammad Bello5;
Paul Joseph Agada6; Ahmad Abubakar Yusuf7; Abdullahi Lawal Rukuna8
2,3,4,5,6,7,8
Co-Authors
1
Department of Computer Science Abubakar Tafawa Balewa University, Bauchi State Nigeria
2
Department of Computer Science National Open University of Nigeria, Nigeria
3
Department of Computer Science Gombe State College of Education & Legel Studies Nafada, Gombe State
Nigeria
4
Department of Computer Science Federal College of Education Technical, Gombe State Nigeria
5
Department of ICT Federal University Gusau, Zamfara State Nigeria
6
Department of Computer Science Plateau State University, Plateau State Nigeria
7
Department of Computer Science Federal Polytechnic Bauchi, Bauchi State Nigeria
8
Department of Computer Science Abubakar Tafawa Balewa University, Bauchi State Nigeria

Publication Date: 2026/01/09

Abstract: The rapid digital transformation of the fashion industry has amplified the need for intelligent tools that can
navigate vast product catalogs and deliver personalized recommendations to consumers. Fashion recommender systems,
leveraging machine learning and deep learning techniques, have emerged as a crucial solution to address challenges such as
choice overload, style compatibility, and changing trends. This review examines the evolution of fashion recommender
systems, from early content-based and collaborative filtering approaches to hybrid, context-aware, and deep learning-based
models capable of processing multimodal data. The methodology involved a structured literature search across major
academic databases, focusing on recent studies that directly address clothing and fashion recommendation. Key findings
reveal that while significant progress has been made in personalization accuracy, diversity, and visual understanding,
persistent limitations remain. These include reliance on small or proprietary datasets, lack of demographic and cultural
diversity, inconsistent evaluation protocols, and domain shifts between curated catalog images and real-world contexts.
Addressing these challenges will require the development of standardized, diverse benchmark datasets, transparent
experimental reporting, and the integration of ethical considerations such as fairness, inclusivity, and privacy. This paper
provides a comprehensive synthesis of existing research, identifies current gaps, and outlines future directions for building
robust, contextually aware, and user-centered fashion recommender systems.

Keywords: Content-Based Filtering, Collaborative Filtering, Hybrid Recommendation Models, Deep Learning, Machine Learning.

How to Cite: Harisu Aliyu; Abdulamalik Abdulsalam; Jamilu Musa; Grace Ojochenemi Emmanuelanorue; Umar Muhammad Bello;
Paul Joseph Agada; Ahmad Abubakar Yusuf; Abdullahi Lawal Rukuna (2025) Machine Learning for Personalized Fashion
Recommendation Systems: A Review. International Journal of Innovative Science and Research Technology,
10(12), 2751-2762. https://doi.org/10.38124/ijisrt/25dec1676

I. INTRODUCTION address this challenge, fashion recommender systems have

emerged as critical tools for improving user experience by
The global fashion industry has witnessed a rapid digital offering personalized product suggestions tailored to
transformation, with online shopping platforms increasingly individual tastes (C.S. et al., 2024).
replacing traditional brick-and-mortar stores (Zhu, 2023). As
e-commerce continues to expand, consumers are exposed to A recommender system is an intelligent algorithmic tool
an overwhelming number of clothing options, making it designed to predict and suggest items that a user may find
difficult to find products that align with their personal interesting based on various forms of data, such as past
preferences, styles, and needs (Madanchian, 2024). To interactions, demographic information, browsing behavior, or

IJISRT25DEC1676 www.ijisrt.com 2751

Volume 10, Issue 12, December – 2025 International Journal of Innovative Science and Research Technology
ISSN No: -2456-2165 https://doi.org/10.38124/ijisrt/25dec1676

product attributes. In the context of fashion, these systems understanding of existing research, technological
aim to simulate the personalized guidance a customer might developments, and practical implementations relevant to
receive from a sales associate, but in a digital environment fashion recommendation. A wide range of academic
(Fayyaz et al., 2020). Fashion recommendation goes beyond databases and scholarly sources were consulted, including
basic item suggestions it must account for complex factors Google Scholar, IEEE Xplore, SpringerLink, ScienceDirect,
such as style compatibility, color coordination, seasonality, the ACM Digital Library, and ResearchGate. These databases
cultural preferences, and current trends (Eldemerdash et al., were selected based on their accessibility, relevance to
2023). computer science and engineering, and their inclusion of
peer-reviewed content.
With the advancement of machine learning and deep
learning techniques, modern fashion recommender systems The search for literature was guided by specific
can now process multimodal data like images, text, and user keywords such as “fashion recommender system,” “clothing
behavior to make accurate, visually-aware, and context- recommendation using machine learning,” “deep learning in
sensitive recommendations. Techniques such as collaborative fashion recommendation,” “image-based recommender
filtering, content-based filtering, hybrid models, and more system,” “hybrid recommender models for fashion,” and
recently, convolutional neural networks and transformers, are “personalized clothing recommendation.” These keywords
increasingly being applied to understand fashion content and were used individually and in various combinations with
user intent (Deldjoo et al., 2024). Boolean operators to refine the results and ensure a focus on
fashion-related applications of recommender systems.
Moreover, the availability of large-scale fashion
datasets such as DeepFashion, Fashion-MNIST, and Amazon The review prioritized publications between the years
Fashion has facilitated the training and evaluation of 2015 and 2025 to capture recent innovations, especially those
sophisticated models. These datasets include high-resolution utilizing machine learning, deep learning, and advanced
images, textual descriptions, categories, and even user artificial intelligence techniques. Only English-language
ratings, making them suitable for a wide range of sources were considered to maintain consistency and
recommendation tasks(Shushi & Adnan M. Abdulazeez, readability, and only studies that directly addressed fashion
2024). recommendation, whether through theoretical models,
algorithmic design, or practical applications, were included in
Despite the progress, fashion recommender systems the final review. Articles that lacked sufficient
still face several challenges, including the cold-start problem methodological detail or were not academically rigorous were
(limited data for new users or items), the subjectivity of excluded.
fashion taste, and the difficulty in capturing user intent in
dynamic contexts such as changing seasons or evolving Relevant information from each selected study was
fashion trends. Additionally, issues of bias, data sparsity, and carefully extracted and analyzed. This included the study’s
lack of diversity in recommendations remain areas of concern objective, the type of recommender system used, the
in existing systems(Joshi et al., 2022). algorithms applied, the datasets employed, the evaluation
metrics, and the main findings. Studies were then grouped
II. REVIEW METHODOLOGY according to the type of recommender approach they used,
such as collaborative filtering, content-based filtering, hybrid
This literature review was conducted using a models, or deep learning techniques. Additionally, attention
qualitative, narrative approach to explore and synthesize was given to how these models were applied within the
current knowledge on fashion recommender systems. The fashion domain, including how they handled challenges like
purpose of this methodology was to ensure a comprehensive visual similarity, user preferences, and seasonal trends.

Fig 1 Article Selection Criteria

IJISRT25DEC1676 www.ijisrt.com 2752

Volume 10, Issue 12, December – 2025 International Journal of Innovative Science and Research Technology
ISSN No: -2456-2165 https://doi.org/10.38124/ijisrt/25dec1676

III. RECOMMENDATION SYSTEM Advanced content-based systems also incorporate

image processing techniques, particularly using computer
Recommender systems are intelligent algorithms vision and convolutional neural networks, to extract visual
designed to analyze user data and suggest items that align features from clothing images. This enables the system to go
with user interests and preferences. They have become an beyond text-based attributes and understand visual aspects
essential component of many online platforms, helping users such as color palettes, garment shape, and style, which are
discover relevant products, content, or services in a highly particularly important in fashion (Shirkhani et al., 2023).
personalized manner (Kang & Wang, 2024). In the context of
fashion, recommender systems assist users in finding One of the main advantages of content-based filtering is
clothing, accessories, and styles that suit their tastes, that it does not require data from other users. This makes it
shopping behavior, and context-specific needs such as effective in cold-start scenarios for users who are new to the
season, event, or cultural background (Deldjoo et al., 2024). system but have provided some initial preferences (Yuan &
Hernandez, 2023). Also, content-based systems can
The primary goal of a recommender system is to reduce continuously refine recommendations as the user interacts
information overload by presenting users with curated with more items.
selections from large product catalogs. These systems utilize
various data sources, such as purchase history, product
features, user demographics, browsing patterns, and even
social interactions, to generate recommendations (Lu et al.,
2015). With the rapid growth of fashion e-commerce and the
increasing demand for personalization, fashion recommender
systems have evolved to incorporate complex algorithms
capable of understanding the visual, textual, and contextual
characteristics of fashion items (Nguyen et al., 2024).

In fashion retail, the effectiveness of recommender

systems directly impacts customer satisfaction, conversion
rates, and brand loyalty. By learning from user interactions
and identifying patterns, these systems enable a more
engaging and efficient shopping experience (Eldemerdash et
al., 2023).

 Types of Recommender System

Several types of recommender systems have been
developed over the years, each with its strengths and
limitations. The most commonly used types include content-
based filtering, collaborative filtering, hybrid systems, and
context-aware approaches.

 Content-Based Filtering Fig 2 Content-Based Filtering, (Mohamed et al., 2019)

Content-based filtering is one of the foundational
approaches in recommender system design. This method  Collaborative Filtering
relies on analyzing the attributes or features of items that a Collaborative filtering models make recommendations
user has previously interacted with such as viewed, liked, based on user-item interaction data, such as ratings, clicks,
rated, or purchased to recommend similar items in the future. and purchase history. These systems identify users with
Rather than relying on the behavior of other users, content- similar behavior and recommend items liked by those users
based systems develop a personalized profile for each user by (Kawasaki & Hasuike, 2017). Collaborative filtering can be
identifying patterns in the content they prefer (Sharad Phalle either user-based or item-based, and is widely adopted due to
& Bhushan, 2024). its ability to provide unexpected or serendipitous
recommendations. However, it struggles with the cold-start
In the fashion domain, this approach involves extracting problem where recommendations cannot be made for new
and analyzing specific product features, such as color, users or items without prior data (Wang et al., 2019).
texture, fabric type, pattern, sleeve length, cut, style (e.g.,
casual, formal, vintage), brand, and category (e.g., dress, Collaborative filtering is one of the most widely used
jacket, shoes). For instance, if a user frequently browses or techniques in recommender systems. Unlike content-based
purchases floral-patterned dresses made of chiffon, the filtering, which relies on item attributes, collaborative
system will recommend other dresses with similar floral filtering makes recommendations by analyzing patterns in
prints or fabric characteristics (Shushi & Adnan M. user behavior and identifying relationships between users and
Abdulazeez, 2024). The strength of content-based filtering items. This method operates under the assumption that users
lies in its ability to make highly personalized suggestions who exhibited similar preferences in the past are likely to
based on detailed item metadata (Chris et al., 2024).

IJISRT25DEC1676 www.ijisrt.com 2753

Volume 10, Issue 12, December – 2025 International Journal of Innovative Science and Research Technology
ISSN No: -2456-2165 https://doi.org/10.38124/ijisrt/25dec1676

share similar tastes in the future (Sharad Phalle & Bhushan, In a typical hybrid framework, the system may
2024). simultaneously analyze user preferences (as in collaborative
filtering) and item features (as in content-based filtering) to
In practice, collaborative filtering systems use user-item generate recommendations. This dual analysis enables the
interaction data such as ratings, product clicks, purchase system to provide suggestions that are not only relevant to a
history, and browsing behavior to identify users with user’s past interactions but also aligned with the
overlapping interests. The system then recommends items characteristics of products they are likely to find appealing
that similar users have liked or interacted with, even if the (Sabiri et al., 2025). For instance, if collaborative filtering
target user has not previously shown interest in those items. identifies that a user shares similar preferences with others
This enables the system to suggest novel or unexpected items, who buy vintage-style clothing, and content-based analysis
thus supporting serendipitous discovery and increasing the confirms the user’s interest in floral patterns, the hybrid
diversity of recommendations (Fareed et al., 2023). model can recommend vintage floral dresses with high
confidence.
There are two primary approaches to collaborative
filtering: user-based and item-based. Hybrid models can be implemented using different
architectural strategies. These include weighted
In user-based collaborative filtering, the system hybridization, where outputs from multiple algorithms are
identifies a group of users whose preferences are similar to combined using a weighted score; switching models, where
the active user and recommends items favored by those peers. the system selects a recommendation technique based on
certain conditions; and model-based hybrids, where features
In item-based collaborative filtering, the system from different models are used within a single unified
identifies similarities between items based on the users who learning framework (Çano & Morisio, 2017). More advanced
interacted with them and recommends items similar to those implementations, particularly in fashion applications, often
the user has already engaged with (Permana, 2024). utilize deep learning architectures to blend collaborative and
content-based signals, enhancing the system's ability to
understand user intent and fashion aesthetics at a deeper level.

In the fashion domain, hybrid recommender systems are

particularly useful due to the multi-dimensional nature of
fashion preferences, which involve not only user-item
interactions but also visual style, seasonality, social trends,
and even emotional appeal (Deldjoo et al., 2024). For
example, a hybrid model can combine visual similarity
metrics from convolutional neural networks with purchase
patterns from user-item interaction matrices to suggest stylish
yet personalized clothing items.

Fig 3 Collaborative Filtering, (Mohamed et al., 2019)

 Hybrid Recommender Systems

Hybrid recommender systems are designed to address
the shortcomings of individual recommendation techniques
by combining two or more algorithms most commonly,
content-based filtering and collaborative filtering. This
integration allows hybrid systems to leverage the strengths of
each approach while mitigating their respective weaknesses,
resulting in more accurate, diverse, and robust
recommendations (Elahi et al., 2023).

IJISRT25DEC1676 www.ijisrt.com 2754

Volume 10, Issue 12, December – 2025 International Journal of Innovative Science and Research Technology
ISSN No: -2456-2165 https://doi.org/10.38124/ijisrt/25dec1676

Fig 4 Hybrid Recommender System: (Khanal et al., 2020)

 Context-Aware Recommender Systems during summer, or recommend formal attire when a user
Context-aware recommender systems represent a more indicates they are attending a professional event. This level of
advanced evolution of traditional recommendation awareness allows for recommendations that are not only
approaches by incorporating situational or contextual factors personalized but also practically relevant (Abugabah et al.,
into the recommendation process. These systems recognize 2020).
that user preferences are not static but often influenced by
external variables such as time of day, location, season, Technically, context-aware systems may use explicit
weather, device type, user mood, or social setting (Mateos & context input, such as user-provided information, or implicit
Bellogín, 2024). context extraction from sensor data like geolocation, calendar
events, weather APIs (Ponce & Abdulrazak, 2022). The
In fashion applications, context plays a particularly contextual data is then integrated into the recommendation
crucial role. Clothing choices are highly dependent on factors algorithm either through pre-filtering (filtering the dataset
such as climate conditions, occasions (e.g., weddings, job based on context before generating recommendations), post-
interviews, parties), cultural norms, and seasonal trends filtering (applying context rules after generating
(Gazzola et al., 2020). A context-aware recommender system recommendations), or contextual modeling, where the
can, for instance, suggest lightweight, breathable clothing context is incorporated directly into the learning model
(Haruna et al., 2017).

IJISRT25DEC1676 www.ijisrt.com 2755

Volume 10, Issue 12, December – 2025 International Journal of Innovative Science and Research Technology
ISSN No: -2456-2165 https://doi.org/10.38124/ijisrt/25dec1676

Fig 5 An Overview of Context-Aware Recommender System Algorithm, (Khan et al., 2020)

 Deep Learning-Based Recommender Systems capture (Deldjoo et al., 2024). By using Convolutional Neural
Deep learning-based recommender systems represent Networks, systems can extract high-level visual features from
the most recent and innovative evolution in the field of clothing images, allowing them to recommend products
recommendation technologies. These systems utilize deep based on visual similarity or style alignment. This is
neural networks to learn complex, non-linear relationships especially useful for tasks such as visual similarity search,
between users and items by processing large-scale, high- outfit generation, and style-based filtering (Hara et al., 2016).
dimensional data. In the context of fashion, deep learning
enables the extraction of intricate visual and textual features Another common deep learning technique used in
from fashion items, the modeling of sequential user behavior, fashion recommendation is the Recurrent Neural Network
and the generation of highly personalized, adaptive (RNN), which excels in modeling sequential and temporal
recommendations (Lee & Kim, 2022). behavior. For instance, an RNN can capture a user's evolving
fashion preferences over time, allowing the system to predict
One of the primary advantages of deep learning in not just what the user likes now, but also what they might
recommender systems is its capacity to handle multi-modal prefer in the future based on recent activity (Yoon & Jang,
data, which is particularly relevant in fashion applications. 2023). Similarly, Long Short-Term Memory (LSTM)
Fashion items are rich in visual content, and their aesthetic networks and Transformer architectures have been employed
appeal often depends on patterns, textures, shapes, and colors to handle more complex sequences and contextual
that traditional recommendation algorithms may fail to dependencies in user behavior (Mswahili & Jeong, 2024).

IJISRT25DEC1676 www.ijisrt.com 2756

Volume 10, Issue 12, December – 2025 International Journal of Innovative Science and Research Technology
ISSN No: -2456-2165 https://doi.org/10.38124/ijisrt/25dec1676

Fig 6 Deep Learning-Based Recommender System Using Cross Convolutional Filters, (Lee & Kim, 2022)

 Application in Fashion Outfit recommendation is a more advanced application

The fashion industry, particularly within the e- where the system suggests combinations of fashion items that
commerce sector, has seen a transformative shift through the go well together (Pralhad Magar & Kondvilkar, 2020). For
integration of recommender systems. These systems have example, it may recommend a pair of shoes and a handbag
become essential tools for enhancing user experience, that match a selected dress. These systems are trained on
improving product discovery, and driving sales through curated outfit datasets or fashion lookbooks and rely on
personalized engagement (Eldemerdash et al., 2023). Unlike compatibility modeling, style recognition, and occasion-
traditional product catalogs, which treat all users equally, based reasoning. Outfit recommendation not only improves
fashion recommender systems provide tailored suggestions cross-selling opportunities but also enhances user satisfaction
based on individual preferences, behavior, context, and style by simplifying the decision-making process.
inclinations. This level of personalization helps address key
challenges in fashion retail, such as choice overload, IV. REVIEW OF RELATED LITERATURE
changing trends, and subjective notions of style (Guan et al.,
2016). The study of fashion recommender systems has gained
significant attention in recent years due to the growing
One of the most common applications of recommender demand for personalized online shopping experiences.
systems in fashion is personalized product recommendation. Researchers have explored various techniques from classical
These systems analyze users' browsing history, purchase filtering methods to cutting-edge deep learning approaches to
behavior, and engagement patterns to suggest items that address the complex nature of fashion recommendation,
match their tastes (Shirkhani et al., 2023). For instance, a user which often involves subjective, visual, and contextual
who frequently views casual streetwear may receive elements.
recommendations aligned with that aesthetic, while another
who shops for formal attire will be shown a different product (Suvarna & Balakrishna, 2024) proposed an enhanced
mix. content-based fashion recommendation system using deep
learning and ensemble techniques. The study focused on
Another key application is visual similarity search, improving recommendation accuracy by leveraging five pre-
where users can find products that are visually similar to an trained convolutional neural networks (CNNs): MobileNet,
image they upload or select. This is particularly useful in DenseNet, Xception, VGG16, and VGG19. These models
fashion, where visual features such as color, cut, print, and were used to extract visual features from fashion product
texture are primary decision-making factors. Powered by images, and their outputs were combined into a deep
convolutional neural networks, these systems can analyze and ensemble classifier. Recommendations were generated based
compare image features to return clothing items with similar on cosine similarity between the feature vectors of the input
visual characteristics. Visual search capabilities are image and the items in the dataset. The system was tested
increasingly being integrated into mobile fashion apps and using two publicly available fashion datasets from Kaggle,
web platforms, enabling users to search for products which included various apparel and shoe images. The
intuitively and efficiently (Jain & Wah, 2022). proposed ensemble model achieved a classification accuracy

IJISRT25DEC1676 www.ijisrt.com 2757

Volume 10, Issue 12, December – 2025 International Journal of Innovative Science and Research Technology
ISSN No: -2456-2165 https://doi.org/10.38124/ijisrt/25dec1676

of 96%, outperforming traditional single-model CNN A personalized fashion recommendation system that
approaches. relies on visual similarity rather than user history was
proposed by (Sridevi et al., 2020). The system uses a
(ASIROGLU et al., 2019) introduced a “smart clothing convolutional neural network, based on ResNet50 and fine-
recommendation system” that requires only a single photo of tuned using transfer learning, to extract visual features from
the user with no prior shopping history to suggest clothing clothing images. Recommendations are generated using
items. The system begins by analyzing the image to detect the cosine similarity and a nearest neighbor algorithm (Annoy
user’s gender using a deep neural network. It then extracts library). The system was trained using the DeepFashion
additional visual features from the photo and searches a dataset and tested on a variety of real-world and internet-
database of garments labeled by gender and other attributes. sourced images. It achieved a high classification accuracy of
The model incorporates transfer learning via a ResNet-50 over 98% during training, demonstrating strong performance
backbone, fine-tuned on a large inventory of fashion images even with a limited dataset. The model effectively addressed
(including data from DeepFashion), allowing it to capture the cold-start problem and produced visually similar
complex visual representations despite limited user input. recommendations based on a single input image, offering a
The study achieve 98% accuracy on color prediction, 86% fast, visually-driven, and scalable solution for fashion
accuracy on gender and cloth's pattern predictions and 75% recommendation.
accuracy on clothing recommendation.
(Sawalkar et al., 2023) developed a content-based
(Yarahmadi Gharaei et al., 2021) developed a content- clothing recommender system using a deep neural network to
based clothing recommender system using a deep neural automatically extract item features such as category and
network that automatically classifies clothing by category and gender directly from images, eliminating manual feature
gender while extracting visual features from images. Trained engineering. The model incorporates demographic
on the Fashion Product Images dataset, the model information, including gender, into the recommendation
outperformed the ResNet-50 baseline with lower loss and process. They compared performance with and without
faster training time. It achieved a recommendation precision demographic features and found that including them led to
of 73.7% based on real user feedback and effectively lower training loss and improved relevance. The system was
addressed the cold-start problem by requiring only item evaluated on an Android-based platform and effectively
images for new product recommendations. The system addressed the cold-start issue, producing novel and
demonstrated high speed, accuracy, and the ability to generate unexpected recommendations for new items. The results
novel and relevant suggestion. demonstrate that their deep learning approach enhances
recommendation accuracy and diversity without requiring
(Lin et al., 2019) developed a content-based clothing historical user data.
recommendation system that integrates personal features
(gender and body height) with clothing attributes (style, SAERS (Semantic Attribute Explainable Recommender
design, color, and texture) to suggest suitable apparel. Using System.) was proposed by (Hou et al., 2019) a fashion
deep learning models like InceptionV3, the system performs recommendation system that provides both accurate and
gender recognition, estimates body height from user photos, explainable suggestions by focusing on fine-grained visual
and identifies clothing attributes from images using the attributes of clothing, such as collar style or heel shape.
FashionAI dataset. These features are combined into a Instead of relying solely on global image features, SAERS
similarity score to match users with appropriate items in a uses a Semantic Extraction Network to identify specific
large clothing gallery. A random generator is also applied to image regions linked to attributes and applies attention
increase recommendation diversity. The system achieved mechanisms to personalize recommendations based on
high attribute recognition accuracy (up to 90.09% for collar individual user preferences. The system is trained with weak
design) and demonstrated strong recommendation supervision and combines both global and attribute-level
performance. In user testing, the system achieved a precision features. Experiments on the Amazon Fashion dataset show
of 71.8% for tops and 64.1% for bottoms when all features that SAERS outperforms existing methods in
were used. recommendation accuracy while also offering interpretable,
attribute-based visual explanations.
(Suvarna & Balakrishna, 2022) proposed a content-
based fashion recommender that uses a custom deep (Sivaranjani et al., 2023) implemented a fashion
convolutional neural network (CNN) to classify a query recommender using a convolutional neural network (CNN)
image and recommend visually similar products. They trained on images to suggest similar apparel and accessories.
trained their model using a subset of nearly 15,000 fashion They divided their dataset into training and testing subsets
images from 12 different categories. The system achieved an and applied CNN for feature extraction, using K-nearest
image classification accuracy of 89.02%, enabling it to neighbors (KNN) to retrieve visually similar items.
generate reliable and precise recommendations. The study Compared to previous systems, their approach delivered
demonstrates that a well-designed deep CNN can effectively results 80.6% faster. The system demonstrated effectiveness
serve as the foundation for a visually-driven, efficient, and in personalized fashion retrieval, handling cold-start
scalable fashion recommendation engine. situations by recommending items based solely on image
input.

IJISRT25DEC1676 www.ijisrt.com 2758

Volume 10, Issue 12, December – 2025 International Journal of Innovative Science and Research Technology
ISSN No: -2456-2165 https://doi.org/10.38124/ijisrt/25dec1676

(Sarojadevi et al., 2023) developed a content-based environmental context in fashion recommendation to

fashion recommendation system that focuses on clothing improve relevance and user satisfaction.
style similarity. Using upper-body and lower-body garment
images along with human model photos, the system extracts (Thakur, 2024) developed an outfit recommendation
visual features through a CNN model based on ResNet-50. It system that integrates user skin tone analysis, real-time
compares these features to a large gallery of clothing images weather data, and a digital wardrobe to suggest personalized
from a Kaggle fashion dataset using similarity matching. The ensembles. A Convolutional Neural Network (CNN) extracts
model effectively handles the cold-start problem by providing visual features from clothing items stored in users’ virtual
outfit suggestions even for users or items with no prior wardrobes. Recommendations are then tailored using k-
interaction data. Nearest Neighbors (kNN), grouping users based on similar
fashion preferences, and refined through decision tree
(Syiam et al., 2023) developed an item-based algorithms. By considering contextual factors like skin tone
collaborative filtering recommender for fashion products and weather, the system provides both stylistically
using purchase data from Rent the Runway. They appropriate and comfortable outfit suggestions while
preprocessed the data by cleaning, normalizing ratings, and promoting sustainable use of existing clothing.
splitting into training and testing sets. The system calculates
item-to-item similarity using adjusted cosine similarity, then (Ramesh & Moh, 2018) presented an outfit
predicts user preferences via a weighted sum of neighbors. recommender system that designs complementary garment
Evaluated using Mean Absolute Error (MAE) and combinations using session-based user interaction data. The
Normalized Discounted Cumulative Gain (NDCG), the model analyzes user sessions series of item clicks and
adjusted cosine method outperformed standard cosine purchases to learn patterns and co-occurrence relationships
similarity, achieving a lower MAE of 0.4424 (versus 0.4815) between fashion items. It leverages a neural network to model
and higher NDCG around 0.9989. This confirms adjusted these sequential interactions, identifying which clothing
cosine as a more accurate metric for generating personalized pieces are likely to be paired together. While the paper
fashion recommendations. focuses primarily on modeling item compatibility rather than
individual preferences, experimental results on real-world e-
(Ye et al., 2023) a novel scene-aware fashion commerce datasets show that the system effectively suggests
recommender system (SAFRS) that recommends outfits based cohesive outfit combinations based on browsing behavior.
on the context or environment (e.g., beach, office, gym). The
system uses attention-based encoders to extract features from V. DATASET USED
both scenes and outfits, then evaluates their compatibility
through a joint scoring function. Trained on a newly curated The performance and generalizability of fashion
dataset with scene-labeled outfits, SAFRS significantly recommender systems are highly influenced by the datasets
outperformed existing models in recommending context- used for their development and evaluation. Table 1
appropriate clothing. The study highlights the importance of summarizes the datasets utilized in the selected studies

Table 1 Summary of Dataset Used

Author Title Dataset
(Suvarna & Balakrishna, Enhanced Content-Based Fashion Recommendation System Fashion Product Images dataset
2024) (Kaggle) + Shoe dataset (Kaggle)
(ASIROGLU et al., 2019) Smart Clothing Recommendation System DeepFashion dataset
(Yarahmadi Gharaei et al., Content-Based Clothing Recommender System using Deep Fashion Product Images dataset
2021) Neural Network (Kaggle)
(Lin et al., 2019) Clothing Recommendation System Based on Visual FashionAI dataset
Information Analytics
(Suvarna & Balakrishna, An Efficient Fashion Recommendation System using a Deep Fashion Product Images dataset
2022) CNN Model (subset, ~15,000 images)
(Sridevi et al., 2020) Personalized Fashion Recommender System with Image- DeepFashion dataset
Based Neural Networks
(Sawalkar et al., 2023) Fashion Recommendation System Custom wardrobe dataset
(Ye et al., 2023) Show Me The Best Outfit for a Certain Scene: A Scene- Scene-labeled outfit dataset
Aware Fashion Recommender System (newly curated by authors)
(Sarojadevi et al., 2023) Fashion Recommender System (FRS): Image Based Engine Kaggle fashion dataset
for Personalized Outfit
(Syiam et al., 2023) Fashion Recommendation System Using Collaborative Rent the Runway dataset
Filtering
(Thakur, 2024) Enhancing Outfit Recommendation with a CNN-kNN Custom-built digital wardrobe
Hybrid Model and Digital Wardrobe Management dataset
(Ramesh & Moh, 2018) Outfit Recommendation Fashion e-commerce dataset from
STYL

IJISRT25DEC1676 www.ijisrt.com 2759

Volume 10, Issue 12, December – 2025 International Journal of Innovative Science and Research Technology
ISSN No: -2456-2165 https://doi.org/10.38124/ijisrt/25dec1676

VI. LIMITATIONS REFERENCES

Despite substantial recent progress, the body of work [1]. Abugabah, A., Cheng, X., & Wang, J. (2020). Learning
reviewed exhibits recurring limitations that constrain the Context-Aware Outfit Recommendation. Symmetry,
generalizability, comparability and practical impact of current 12(6), 873. https://doi.org/10.3390/sym12060873
fashion recommender systems. A dominant limitation [2]. ASIROGLU, B., ATALAY, M. I., BALKAYA, A.,
concerns the data. Many studies rely on small or narrowly TUZUNKAN, E., Dagtekin, M., & ENSARI, T.
curated datasets, or on proprietary databases and custom (2019). Smart Clothing Recommendation System with
“digital wardrobe” collections that were collected for a single Deep Learning. 2019 3rd International Symposium on
experiment. Such datasets frequently suffer from class Multidisciplinary Studies and Innovative Technologies
imbalance, limited demographic and cultural diversity, and (ISMSIT), 1–4.
inconsistent or noisy annotations. The disparity between https://doi.org/10.1109/ISMSIT.2019.8932738
studio/catalog images and in the wild photos like street, user [3]. Çano, E., & Morisio, M. (2017). Hybrid recommender
uploads, social media creates a domain shift that models systems: A systematic literature review. Intelligent
trained on catalog data often cannot bridge. The common Data Analysis, 21(6), 1487–1524.
practice of using different, non-standard splits and https://doi.org/10.3233/IDA-163209
undocumented preprocessing steps further prevents fair [4]. Chris, E., Joe, E., & Grace, A. (2024). Metadata in
comparison between methods and hinders reproducibility. Recommender Systems.
https://www.researchgate.net/publication/385774403
These constraints underscore the need for larger, more [5]. C.S., D., R, K., & N, S. (2024). Enhancing E-
diverse, and standardized datasets, along with transparent and Commerce with Personalized Product
reproducible experimental protocols, to ensure the Recommendations.
advancement and real world applicability of fashion https://doi.org/10.20944/preprints202410.2506.v1
recommender systems. [6]. Deldjoo, Y., Nazary, F., Ramisa, A., McAuley, J.,
Pellegrini, G., Bellogin, A., & Noia, T. Di. (2024). A
VII. CONCLUSION Review of Modern Fashion Recommender Systems.
ACM Computing Surveys, 56(4), 1–37.
This review has examined the current state of fashion https://doi.org/10.1145/3624733
recommender systems, outlining their methodologies, [7]. Elahi, M., Khosh Kholgh, D., Kiarostami, M. S.,
datasets, and applications, while highlighting recurring Oussalah, M., & Saghari, S. (2023). Hybrid
limitations. The literature reveals a clear evolution from basic recommendation by incorporating the sentiment of
content-based and collaborative filtering techniques to more product reviews. Information Sciences, 625, 738–756.
sophisticated hybrid and context-aware models that integrate https://doi.org/10.1016/j.ins.2023.01.051
visual, textual, and contextual signals. These advances have [8]. Eldemerdash, D., AL shikh, K., & Abou-Ghali, M.
improved personalization, recommendation diversity, and (2023). Fashion Recommendation System and its
adaptability to varying user needs and scenarios. However, Impact on Consumers’ Purchase Decision Making.
persistent challenges remain. Issues such as limited and non- International Design Journal, 13(1), 195–206.
standardized datasets, lack of demographic diversity, https://doi.org/10.21608/idj.2023.276188
inadequate modeling of contextual factors, and inconsistent [9]. Fareed, A., Hassan, S., Belhaouari, S. B., & Halim, Z.
evaluation protocols hinder both the reproducibility and the (2023). A collaborative filtering recommendation
real-world applicability of these systems. Ethical framework utilizing social networks. Machine
considerations, including fairness, inclusivity, and user Learning with Applications, 14, 100495.
privacy, are still underexplored in most works, leaving gaps https://doi.org/10.1016/j.mlwa.2023.100495
between academic research and commercial deployment. [10]. Fayyaz, Z., Ebrahimian, M., Nawara, D., Ibrahim, A.,
Going forward, the field would benefit from collaborative & Kashef, R. (2020). Recommendation Systems:
efforts to establish standardized, diverse benchmark datasets, Algorithms, Challenges, Metrics, and Business
transparent experimental reporting, and evaluation Opportunities. Applied Sciences, 10(21), 7748.
frameworks that incorporate both offline metrics and real- https://doi.org/10.3390/app10217748
world user engagement. Furthermore, incorporating [11]. Gazzola, P., Pavione, E., Pezzetti, R., & Grechi, D.
multimodal data, addressing cold-start problems, and (2020). Trends in the Fashion Industry. The Perception
embedding fairness and privacy mechanisms into system of Sustainability and Circular Economy: A
design will be crucial. By tackling these challenges, future Gender/Generation Quantitative Approach.
fashion recommender systems can evolve into robust, Sustainability, 12(7), 2809.
inclusive, and contextually aware solutions that meaningfully https://doi.org/10.3390/su12072809
enhance user experience and drive innovation in the fashion [12]. Guan, C., Qin, S., Ling, W., & Ding, G. (2016).
industry. Apparel recommendation system evolution: an
empirical review. International Journal of Clothing
Science and Technology, 28(6), 854–879.
https://doi.org/10.1108/IJCST-09-2015-0100
[13]. Hara, K., Jagadeesh, V., & Piramuthu, R. (2016).
Fashion apparel detection: The role of deep

IJISRT25DEC1676 www.ijisrt.com 2760

Volume 10, Issue 12, December – 2025 International Journal of Innovative Science and Research Technology
ISSN No: -2456-2165 https://doi.org/10.38124/ijisrt/25dec1676

convolutional neural network and pose-dependent [25]. Madanchian, M. (2024). The Impact of Artificial
priors. 2016 IEEE Winter Conference on Applications Intelligence Marketing on E-Commerce Sales.
of Computer Vision (WACV), 1–9. Systems, 12(10), 429.
https://doi.org/10.1109/WACV.2016.7477611 https://doi.org/10.3390/systems12100429
[14]. Haruna, K., Akmar Ismail, M., Suhendroyono, S., [26]. Mateos, P., & Bellogín, A. (2024). A systematic
Damiasih, D., Pierewan, A., Chiroma, H., & Herawan, literature review of recent advances on context-aware
T. (2017). Context-Aware Recommender System: A recommender systems. Artificial Intelligence Review,
Review of Recent Developmental Process and Future 58(1), 20. https://doi.org/10.1007/s10462-024-10939-
Research Direction. Applied Sciences, 7(12), 1211. 4
https://doi.org/10.3390/app7121211 [27]. Mohamed, M. H., Khafagy, M. H., & Ibrahim, M. H.
[15]. Hou, M., Wu, L., Chen, E., Li, Z., Zheng, V. W., & Liu, (2019). Recommender Systems Challenges and
Q. (2019). Explainable Fashion Recommendation: A Solutions Survey. 2019 International Conference on
Semantic Attribute Region Guided Approach. Innovative Trends in Computer Engineering (ITCE),
http://vision.cs.utexas.edu/projects/finegrained/utzap5 149–155. https://doi.org/10.1109/ITCE.2019.8646645
0k/ [28]. Mswahili, M. E., & Jeong, Y.-S. (2024). Transformer-
[16]. Jain, V., & Wah, C. (2022). Computer Vision in based models for chemical SMILES representation: A
Fashion Trend Analysis and Applications. Journal of comprehensive literature review. Heliyon, 10(20),
Student Research, 11(1). e39038.
https://doi.org/10.47611/jsrhs.v11i1.2464 https://doi.org/10.1016/j.heliyon.2024.e39038
[17]. Joshi, A., Wong, C. L., Marinho de Oliveira, D., Zafari, [29]. Nguyen, D.-N., Nguyen, V.-H., Trinh, T., Ho, T., & Le,
F., Mourao, F., Ribas, S., & Pandey, S. (2022). H.-S. (2024). A personalized product recommendation
Imbalanced Data Sparsity as a Source of Unfair Bias model in e-commerce based on retrieval strategy.
in Collaborative Filtering. Proceedings of the 16th Journal of Open Innovation: Technology, Market, and
ACM Conference on Recommender Systems, 531– Complexity, 10(2), 100303.
533. https://doi.org/10.1145/3523227.3547404 https://doi.org/10.1016/j.joitmc.2024.100303
[18]. Kang, L., & Wang, Y. (2024). Efficient and accurate [30]. Permana, K. E. (2024). Comparison of User Based and
personalized product recommendations through Item Based Collaborative Filtering in Restaurant
frequent item set mining fusion algorithm. Heliyon, Recommendation System. Mathematical Modelling of
10(3), e25044. Engineering Problems, 11(7), 1922–1928.
https://doi.org/10.1016/j.heliyon.2024.e25044 https://doi.org/10.18280/mmep.110723
[19]. Kawasaki, M., & Hasuike, T. (2017). A [31]. Ponce, V., & Abdulrazak, B. (2022). Context-Aware
recommendation system by collaborative filtering End-User Development Review. Applied Sciences,
including information and characteristics on users and 12(1), 479. https://doi.org/10.3390/app12010479
items. 2017 IEEE Symposium Series on [32]. Pralhad Magar, S., & Kondvilkar, S. (2020). MODA-
Computational Intelligence (SSCI), 1–8. A fashion community and outfit recommender
https://doi.org/10.1109/SSCI.2017.8280983 Android application.
[20]. Khan, A., Ahmad, A., Rahman, A. U., & Alkhalil, A. https://www.researchgate.net/publication/342154496
(2020). A Mobile Cloud Framework for Context- [33]. Ramesh, N., & Moh, T.-S. (2018). Outfit
Aware and Portable Recommender System for Smart Recommender System. 2018 IEEE/ACM International
Markets (pp. 283–309). https://doi.org/10.1007/978-3- Conference on Advances in Social Networks Analysis
030-13705-2_12 and Mining (ASONAM), 903–910.
[21]. Khanal, S. S., Prasad, P. W. C., Alsadoon, A., & Maag, https://doi.org/10.1109/ASONAM.2018.8508656
A. (2020). A systematic review: machine learning [34]. Sabiri, B., Khtira, A., El Asri, B., & Rhanoui, M.
based recommendation systems for e-learning. (2025). Hybrid Quality-Based Recommender
Education and Information Technologies, 25(4), Systems: A Systematic Literature Review. Journal of
2635–2664. https://doi.org/10.1007/s10639-019- Imaging, 11(1), 12.
10063-9 https://doi.org/10.3390/jimaging11010012
[22]. Lee, S., & Kim, D. (2022). Deep learning based [35]. Sarojadevi, H., Bhat, V., Palekar, R., & Neha, K.
recommender system using cross convolutional filters. (2023). Fashion Recommender System (FRS): Image
Information Sciences, 592, 112–122. Based Engine for Personalized Outfit. Asian Journal of
https://doi.org/10.1016/j.ins.2022.01.033 Computer Science and Technology, 12(1), 21–24.
[23]. Lin, Y.-R., Su, W.-H., Lin, C.-H., Wu, B.-F., Lin, C.- https://doi.org/10.51983/ajcst-2023.12.1.3569
H., Yang, H.-Y., & Chen, M.-Y. (2019). Clothing [36]. Sawalkar, P., Udar, B., Bobade, A., Gupta, N., &
Recommendation System based on Visual Information Kahar, S. (2023). FASHION RECOMMENDATION
Analytics. 2019 International Automatic Control SYSTEM. JETIR. www.jetir.orgm650
Conference (CACS), 1–6. [37]. Sharad Phalle, Ms. T., & Bhushan, Prof. S. (2024).
https://doi.org/10.1109/CACS47674.2019.9024361 Content Based Filtering And Collaborative Filtering:
[24]. Lu, J., Wu, D., Mao, M., Wang, W., & Zhang, G. A Comparative Study. Journal of Advanced Zoology,
(2015). Recommender system application 96–100. https://doi.org/10.53555/jaz.v45iS4.4158
developments: A survey. Decision Support Systems, [38]. Shirkhani, S., Mokayed, H., Saini, R., & Chai, H. Y.
74, 12–32. https://doi.org/10.1016/j.dss.2015.03.008 (2023). Study of AI-Driven Fashion Recommender

IJISRT25DEC1676 www.ijisrt.com 2761

Volume 10, Issue 12, December – 2025 International Journal of Innovative Science and Research Technology
ISSN No: -2456-2165 https://doi.org/10.38124/ijisrt/25dec1676

Systems. SN Computer Science, 4(5), 514. 2023, 1172–1180.

https://doi.org/10.1007/s42979-023-01932-9 https://doi.org/10.1145/3543507.3583435
[39]. Shushi, A., & Adnan M. Abdulazeez. (2024). Fashion [49]. Yoon, J. H., & Jang, B. (2023). Evolution of Deep
Design Classification Based on Machine Learning and Learning-Based Sequential Recommender Systems:
Deep Learning Algorithms: A Review. Indonesian From Current Trends to New Perspectives. IEEE
Journal of Computer Science, 13(3). Access, 11, 54265–54279.
https://doi.org/10.33022/ijcs.v13i3.3980 https://doi.org/10.1109/ACCESS.2023.3281981
[40]. Sivaranjani, L., Rachamadugu, S. K., Reddy, B. V. S., [50]. Yuan, H., & Hernandez, A. A. (2023). User Cold Start
A, B. R., Sakthivel, M., & Depuru, S. (2023). Fashion Problem in Recommendation Systems: A Systematic
Recommendation System Using Machine Learning. Review. IEEE Access, 11, 136958–136977.
2023 4th International Conference on Smart https://doi.org/10.1109/ACCESS.2023.3338705
Electronics and Communication (ICOSEC), 1367– [51]. Zhu, Z. (2023). The Impact of Digital Transformation
1374. on the Fashion Industry in the Post-Pandemic Era.
https://doi.org/10.1109/ICOSEC58147.2023.1027596 BCP Business & Management, 44, 696–705.
7 https://doi.org/10.54691/bcpbm.v44i.4920
[41]. Sridevi, M., ManikyaArun, N., Sheshikala, M., &
Sudarshan, E. (2020). Personalized fashion
recommender system with image based neural
networks. IOP Conference Series: Materials Science
and Engineering, 981, 022073.
https://doi.org/10.1088/1757-899X/981/2/022073
[42]. Suvarna, B., & Balakrishna, S. (2022). An Efficient
Fashion Recommendation System using a Deep CNN
Model. 2022 International Conference on Automation,
Computing and Renewable Systems (ICACRS), 1179–
1183.
https://doi.org/10.1109/ICACRS55517.2022.1002906
3
[43]. Suvarna, B., & Balakrishna, S. (2024). Enhanced
content-based fashion recommendation system
through deep ensemble classifier with transfer
learning. Fashion and Textiles, 11(1), 24.
https://doi.org/10.1186/s40691-024-00382-y
[44]. Syiam, M. K., Wibowo, A. T., & Setiawan, E. B.
(2023). Fashion Recommendation System using
Collaborative Filtering. Building of Informatics,
Technology and Science (BITS), 5(2).
https://doi.org/10.47065/bits.v5i2.3690
[45]. Thakur, S. (2024). Enhancing Outfit Recommendation
with a CNN-kNN Hybrid Model and Digital Wardrobe
Management. International Journal for Research in
Applied Science and Engineering Technology, 12(4),
2463–2467.
https://doi.org/10.22214/ijraset.2024.60378
[46]. Wang, C.-D., Deng, Z.-H., Lai, J.-H., & Yu, P. S.
(2019). Serendipitous Recommendation in E-
Commerce Using Innovator-Based Collaborative
Filtering. IEEE Transactions on Cybernetics, 49(7),
2678–2692.
https://doi.org/10.1109/TCYB.2018.2841924
[47]. Yarahmadi Gharaei, N., Dadkhah, C., & Daryoush, L.
(2021). Content-based Clothing Recommender
System using Deep Neural Network. 2021 26th
International Computer Conference, Computer
Society of Iran (CSICC), 1–6.
https://doi.org/10.1109/CSICC52343.2021.9420544
[48]. Ye, T., Hu, L., Zhang, Q., Lai, Z. Y., Naseem, U., &
Liu, D. D. (2023). Show Me The Best Outfit for A
Certain Scene: A Scene-aware Fashion Recommender
System. Proceedings of the ACM Web Conference

IJISRT25DEC1676 www.ijisrt.com 2762