Volume 9, Issue 4, April – 2024 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24APR1249

Recommender System in E-Commerce

Abhinav Sharma1; Preksha Agrawal2; Surendra Kumar Keshari3
Information Technology KIET Group of Institutions Delhi-NCR, Ghaziabad, India

Abstract:- In the realm of e-commerce, recommendation This project aims to leverage advanced technology,
systems play a pivotal role in guiding users towards particularly graph databases such as Neo4j, to augment the
relevant products. However, existing systems often capabilities of these systems. Graph databases offer a unique
grapple with inefficiencies in handling large datasets and approach to organizing and querying interconnected data,
fail to deliver personalized recommendations tailored to offering a promising avenue for developing more
individual preferences. Addressing these challenges, the sophisticated and personalized recommendation algorithms
study introduces an innovative approach leveraging [2].
graph databases to enhance the performance of e-
commerce recommendation systems. Through At the core of this study lies the objective of conducting
comprehensive analysis, the study delves into four a comprehensive evaluation and comparison of Neo4j and
critical aspects: database comparison, user exploration MySQL databases within the context of an e-commerce
frequency across product categories, the diversity of recommender system. By scrutinizing their performance in
available category types, and user browsing history handling query parameters and response times, the aim is to
analysis. This investigation unveils Neo4j's superior determine the most suitable database solution for managing
efficiency over MySQL in managing extensive datasets, extensive datasets, thereby ensuring timely and accurate
laying the groundwork for more robust recommendation information retrieval. Furthermore, study delves into
engines. By scrutinizing user behaviour patterns, the specific metrics pivotal to the effectiveness of the discussed
recommender system predicts preferences with system, including the frequency of user visits to different
precision, promising a tailored and gratifying shopping grouping types, the diversity of available grouping types, and
experience for users. Moreover, with support for a patterns in user traversal history [3]. Fig 1 discuss about how
diverse array of category types, users gain flexibility in the recommendation works on abstract level. This Fig 1. tell
exploring products based on varied criteria, addressing more about factors that are used in discussed
a crucial need in the market for personalized shopping recommendation system.
experiences. Leveraging insights gleaned from user
browsing history, the system delivers refined Observations seeks to bridge the gap between
recommendations, poised to elevate user satisfaction and traditional recommendation systems and the evolving needs
engagement within the competitive landscape of e- of online shoppers. Through this endeavour, the aspiration to
commerce. In conclusion, the study highlights the enhance the overall shopping experience, ultimately driving
significance of recommendation systems in enhancing the user engagement and satisfaction in the ever-expanding
e- commerce experience. By leveraging graph databases, landscape of e-commerce.
particularly Neo4j, over traditional systems like MySQL,
significant improvements in managing extensive datasets
are demonstrated.

Keywords:- E-commerce, Recommendation Systems, Neo4j,

Personalized Recommendations, User Engagement, Data
Analysis.

I. INTRODUCTION

In the ever-evolving realm of online retail,

recommendation systems play a central role in shaping user
experiences and influencing purchasing behaviours. As e-
commerce platforms strive to deliver tailored and
compelling content, there exists a pressing need for novel
solutions to enhance the efficiency and effectiveness of
recommendation engines [1]. This research endeavours to
tackle the inherent challenges faced by traditional
recommendation systems, with a keen focus on enhancing
user satisfaction and refining the precision of product
suggestions. Fig 1: Conceptual Diagram for Recommendation System

IJISRT24APR1249 www.ijisrt.com 2576

Volume 9, Issue 4, April – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24APR1249

II. LITERATURE REVIEW B. Enhanced User Profiling through Tagging Behaviour

Analysis
The section discusses the proposed recommender To improve recommendation accuracy, De Gemmis et
engine encompasses a comprehensive understanding of al. [12] emphasized the importance of user tagging
existing recommendation systems, particularly collaborative behaviour in constructing more accurate user profiles.
filtering methods and the challenges they face. Collaborative Gedikli et al. [13] expanded on this notion, proposing
filtering (CF) methods, including memory-based and model- methodologies to exploit context- specific tag preferences
based approaches, have been extensively studied and utilized within recommendation processes.
in various online platforms to provide personalized
recommendations based on past user- item interactions. Some attempts by Du et al.[14] aimed to reduce
misinterpretation of user preferences derived solely from
A. Collaborative Filtering (CF) Recommendation Methods tags.
The recommendation systems field encompasses
Collaborative Filtering (CF), Content-Based Filtering C. Challenges and Innovations in Addressing Data Sparsity
(CBF), and hybrid methods. CF strategies focus on through Model-Based Approaches
computing similarities between users and items using Matrix Factorization (MF) methods, a prominent
distance measures, with variations between user-based and model-based technique, aim to predict ratings by
item-based approaches. decomposing the user-item matrix into latent feature vectors.
SVD++ and probabilistic matrix decomposition models by
Recommender systems commonly leverage two CF Koren et al. [15] initially struggled with data sparsity,
techniques: memory-based and model-based methods as limiting their ability to extract comprehensive user-item
mentioned by Breese et al. [4] feature information.

Clustering techniques are instrumental in model- based Subsequently, Enrich et al. [16] and Bao et al. [17]
CF approaches, discussed by Sarwar et al. [5]. Recent studies proposed advanced algorithms integrating tag information to
explore model-based clustering approaches to enhance improve rating predictions and alleviate overfitting issues in
efficiency. matrix decomposition.

Traditionally, data collection in business was primarily D. Leveraging Temporal and Tagging Data Integration for
for specific purposes. However, the realization of data's Enhanced Recommendations
value as a substantial asset has transformed its treatment. For Recent studies, such as Wang et al. [18] in TV program
instance, the IBM and Twitter partnership on data analytics recommendations, have explored temporal information to
aimed at selling analytical information to corporate clients, enhance group recommendations. This study aims to delve
enabling real-time decision-making [6]. deeper into integrating and exploiting additional tagging and
temporal data sources, aiming to better understand user tag
Big data represents a recent evolution in decision preferences and deliver more precise recommendations.
support data management [7], significantly impacting
various business domains, including Customer Relationship E. Utilizing Graph Databases in recommendations
Management (CRM), Enterprise Resource Planning (ERP), The proposed recommender engine represents a
and Supply Chain Management (SCM). paradigm shift in recommendation systems, leveraging
graph databases to overcome existing limitations. Unlike
The variety aspect involves handling various data conventional methods, it amalgamates graph databases'
types, including structured and unstructured data, from strengths to address various loopholes prevalent in the current
diverse channels like audio, video, images, and location data landscape.
(Google Maps, web pages, text) [8].
One primary shortcoming of memory-based
Customer Relationship Management or CRM serves as collaborative filtering methods is their susceptibility to
a pivotal tool and strategy for managing customer sparsity issues, especially in large-scale applications.
interactions, encompassing sales, marketing, and customer Leveraging graph structures allows for meaningful
service activities [9]. representation of user-item interactions, addressing the
sparsity problem.
The emergence of Web 2.0 has revolutionized
collaboration platforms beyond traditional email and The engine utilizes graph-based models to enhance
information retrieval [10]. Social networks embedded within recommendation accuracy and scalability.
Web 2.0 have transformed organizations by enabling peer-to-
peer collaboration and real-time communication [11]. Furthermore, the engine's use of graph databases
However, this rapid growth in Web 2.0 has necessitated a enables the hybridization of recommendation techniques
shift in CRM techniques, leading to the advent of Social within a unified structure, offering a more robust
CRM (CRM 2.0). recommendation mechanism to users.

IJISRT24APR1249 www.ijisrt.com 2577

Volume 9, Issue 4, April – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24APR1249

Its architecture allows for real-time data processing, In Equation (2), searched Product[i] and product[i] are
leveraging the speed of graph databases for adaptive learning, the binary values (0 or 1) for the ith grouping type in the
ensuring recommendations remain relevant and up-to-date. binary arrays of the searched product and another product,
respectively.
III. METHODOLOGY
-n is the total number of grouping types.
The methodology encompasses defining user and
tracker nodes, collecting tracker node values, creating and This formula computes the cosine similarity between
updating relationships based on user interactions, storing the binary arrays of the searched product and another
grouping information, generating binary arrays, calculating product, providing a measure of how similar their grouping
cosine similarity, sorting products, and presenting top type preferences are. The resulting similarity score is used in
recommendations. By leveraging collaborative filtering sorting and ranking products to present personalized
concepts and frequency-based considerations, the system recommendations.
provides personalized and accurate recommendations
tailored to individual user preferences, ensuring  Present Top Recommendations: The formula involves
comprehensive ranking and presentation of top presenting the top recommendations to the user based on
recommendations. the sorted products. The function Present Top
Recommendations (sorted Products, top N) incorporates
 Define User and Tracker Nodes: The creation of user collaborative filtering and frequency-based
and tracker nodes is a foundational step, initializing the considerations to deliver personalized and accurate
graph to represent users and tracker attributes. recommendations.
 Collect Tracker Node Values: The function Collect
Tracker Node Values (products, tracker Nodes) gathers The Flowchart discuss the methodology and flow of
values associated with tracker nodes for each control. Fig 2. discusses about the flow of algorithm and data
product, facilitating a detailed understanding of user in recommender system.
preferences based on tracking attributes.
 Store Grouping Information: The part involves
storing grouping information based on frequency
considerations. The function Store Grouping Information
(user Nodes, tracker Nodes, interactions) identifies and
stores the most frequently visited grouping types and
values, contributing to a richer understanding of user
preferences.
 Generate Binary Array: The part involves generating
binary arrays based on the frequency of grouping
values. The function Generate Binary Array(products,
grouping Info) emphasizes the importance of frequently
visited values in user preferences by representing them as
binary arrays.
 Calculate Cosine Similarity: The formula involves
applying collaborative filtering concepts in the cosine
similarity calculation.

The cosine similarity formula is used in the Calculate

Cosine Similarity function to measure the similarity between
the searched product and other products. Here's the cosine
similarity formula in Equation (1):

Given two vectors A and B, the cosine similarity

(cos(θ)) is calculated as:

In the context of the recommender system

methodology:

Fig 2: Flowchart of Recommender System

IJISRT24APR1249 www.ijisrt.com 2578

 Volume 9, Issue 4, April – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24APR1249

IV. PROPOSED PRODUCT RECOMMENDATION C. Attribute and Feature Integration

USING GRAPH DATABASE With the foundational structure in place, the project
delves into the creation of attributes and their associations
To generate accurate recommendations, the products and with respective categories. For instance, create an attribute
their grouping data should be represented in graph database named "Brand" for the "Apparel" category:
in Neo4j. After the data, has been represented in graph
database, customers can get product recommendations using CREATE (:Product{name:"Nike Air Max", id:101,
cypher queries. The Fig 3. and Fig 4. shows outputs of cypher brand:'Nike'})
queries. MATCH (a:Product),(b:Value) WHERE a.id=101 AND
a.brand=b.name CREATE (a)- [:Related{type:'Brand'}]-
A. Dataset Description >(b) RETURN b
In the context of their e-commerce product
recommendation project, data organization occurs within a D. Attribute Value Creation
graph database framework. The dataset construction follows The dataset incorporates attribute values that can be
a sequence of distinct steps, each characterized by specific universally accessible across multiple categories and
queries and relational structures. attributes. Establish relationships between attributes and their
corresponding values, allowing attributes like "Brand" to be
 Grouping Data: For the creation of groupings based on associated with specific values:
the store's grouping data, utilize API calls to receive the
necessary information. MATCH (a:Attribute),(b:Value) WHERE
 Product Creation: Following the establishment of a.isAttributeOf="Footwear" AND a.name="Brand" AND
groupings, products are created in alignment with the b.name="Adidas" CREATE (a)-[:ItemValue]-
grouping and grouping type data associated with each >(b) RETURN a,b
product.
 Cosine Similarity: Cosine similarity plays a pivotal role
in the recommendation system. The similarity between
products is calculated using feature arrays (val and val1).
Here's a working query demonstrating the computation of
cosine similarity:
 WITH gds.similarity.cosine(val, val1) as res, p2 as p2
RETURN p2.code, p2.name, p2.thumbnail, res ORDER
BY res DESC LIMIT 10.

B. Dataset Structure Establishment

Project begins by creating a structured database for
housing product-related data, employing a graph database
model. Initiate this process by establishing a main node in
database with predefined categories, including "Apparel,"
"Footwear," "Personal Care," and "Sporting Goods".
Subsequently, Forge relationships between these categories
and the main database to signify their inclusion within the
overarching dataset structure.

MATCH (a:Category),(b:Database) WHERE NOT

EXISTS((b)-[:Include]->(a)) CREATE (b)-[:Include]->(a) Fig 4. Output for Relationships between
RETURN a,b Attributes and their Values

E. Product Creation and Attribute Linkage:

To complete the dataset, product nodes are created and
directly relate attributes like "Brand" to their corresponding
values. For instance, create a product named "Nike Air Max"
and establish its brand as "Nike".

CREATE (:Product{name:"Nike Air Max", id:101,

brand:'Nike'})
MATCH (a:Product),(b:Value) WHERE a.id=101 AND
a.brand=b.name CREATE (a)- [:Related{type:'Brand'}]-
>(b) RETURN b

Fig 3: Output for Relationship between Categories

IJISRT24APR1249 www.ijisrt.com 2579

Volume 9, Issue 4, April – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24APR1249

F. Attribute Value Count system's behaviour concerning the database comparison,

Finally, the dataset is leveraged to determine the count frequency of user visits, the diversity of grouping types
of attribute values linked to a specific category, relying on available within the system, and the history of user traversal
the product's relation type. For instance, the counting process patterns to ascertain its ability to deliver tailored and effective
encompasses values associated with the "Brand" attribute for recommendations.
the product "Nike Air Max."
MATCH (a:Product)-[r:Related]->(b:Value)<- A. Database Efficiency and Comparison with traditional
[v:ItemValue]-(c:Attribute)-[:ItemValue]->(f:Value) SQL Database
WHERE a.id=101 AND b.name=a.brand AND The aim is to see which database, Neo4j or MySQL,
c.name=r.type RETURN COUNT(f) + 1. worked better for e-commerce project. Same Questions were
asked to both databases five times and averaged how long it
This comprehensive dataset construction approach took to get answers. The main goal is to find out which
facilitates the creation of a rich and interconnected database, database could handle a lot of data and provide quick
paving the way for effective product recommendations answers.
within the e-commerce platform.
For e-commerce work, store important information in
V. RESULT AND DISCUSSION Neo4j, a special type of database. The study used it to ask
questions and get recommendations for products. At the
In this investigation, the research focuses on the same time, store the same information in MySQL to compare
comprehensive evaluation of the recommender system, how fast both databases worked.The results, shown in Table
investigating its performance across key parameters 1, highlight Neo4j's quicker performance compared to
essential for efficient recommendation generation in e- MySQL. Times are measured in milliseconds (ms).
commerce platforms. Specifically, examine and compare the

Table 1: Comparison between Neo4j and MySQL

Metric Neo4j MySQL
Query Complexity (Out of 10) 4.5 7.2
Indexing Efficiency (Out of 10) 8.2 5.9
Response Time (ms) 12.1 18.5
Scalability Excellent Good

Neo4j's superior indexing efficiency over MySQL is query language position it as a more effective solution for
deeply rooted in its native graph structure, purpose- built to scenarios where query complexity is a critical factor.
excel in graph databases. Its support for native full-text
indexing and composite indexes further enhances efficiency, B. Frequency of Time Each Grouping Type is Visited:
offering a compelling advantage over MySQL in scenarios In the context of online platforms, e-commerce, and
prioritizing relationships, adaptability, and tailored indexing search engines, the term "frequency" refers to the number
strategies. of times a product or grouping has been searched. This
metric is used in analytics and business intelligence to
Neo4j outshines MySQL in query complexity due to its understand user behaviour.
native graph database architecture. Neo4j's strength lies in
traversing relationships, making it inherently suited for The frequency of a product or grouping is measured by
complex graph queries. In contrast, MySQL, may require how often users enter specific search queries related to that
complex joins and subqueries to achieve similar graph-based product or category. This information can be gathered
operations, resulting in increased query complexity. through search logs and analytics tools provided by search
engines or e-commerce platforms. High search frequency for
The label-based indexing in Neo4j contributes a particular product or grouping may indicate strong user
significantly to its favorable query complexity. This indexing interest and demand. Products with high search frequency
mechanism, designed for graph-oriented searches, aligns are often considered popular, and businesses may use this
seamlessly with the structure of the graph database, information to optimise their inventory, marketing
simplifying the formulation of queries based on node labels. campaigns, and overall product offerings.
MySQL, with its traditional B-tree indexes, may face
challenges in achieving the same level of simplicity in Users can leverage information about search frequency
queries involving relationships, potentially leading to higher to plan targeted marketing campaigns and promotions.
query complexity. Promoting products that have a high search frequency can be
a strategic way to capture user attention and drive sales.
Moreover, Neo4j's Cypher query language supports
graph traversal operations with a natural syntax, making it For a user, who often visits website and searches for
easier to express and understand queries that involve various products. So system may save frequency for a
navigating through nodes and relationships. In essence, grouping type here "Brand" as represented in Table II.
Neo4j's native support for graph queries and its optimized

IJISRT24APR1249 www.ijisrt.com 2580

Volume 9, Issue 4, April – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24APR1249

Table 2: Working of Frequency Metric in System The combination of high visit frequency in specific
GROUPING VALUE FREQUENCY groupings, a diverse range of grouping types, and a deep
Casio 1 understanding of user traversal patterns positions the
Titan 3 recommender system to deliver tailored and effective
FastTrack 0 recommendations. Leveraging these crucial parameters
Esprit 1 enhances the system's ability to meet user preferences and
Timex 0 improve overall user satisfaction.

So it can be easily predicted that Titan is most visited by VI. CONCLUSION

user
The shortcomings of conventional recommendation
C. Grouping Types in the System: systems, including slow execution times and limited
Grouping type refers to the manner in which elements personalization. By leveraging innovative database
or entities are categorised, organised, or classified based on technologies such as Neo4j, they've demonstrated significant
shared characteristics, properties, or criteria. grouping types improvements in handling large datasets, addressing a critical
play a pivotal role in organising and managing information bottleneck in system efficiency. Moreover, insights gleaned
efficiently. There are several types of grouping. The from analysing user visit frequency, grouping type diversity,
commonly used grouping type are color, gender, age, brand, and historical traversal patterns have enabled the delivery of
etc. tailored suggestions, meeting individual preferences and
enhancing overall satisfaction in online shopping
Each grouping type serves a unique purpose, providing experiences.
a flexible framework for structuring data and enabling
efficient retrieval and analysis, tailored to the specific needs Looking ahead, the integration of efficient databases
of the problem domain. The choice of grouping type is a and insightful metric analysis offers promising avenues for
critical design decision, influencing the ease of data further advancements in personalized recommendation
manipulation, search, and overall system performance. systems. This underscores the importance of aligning
technology with user needs, paving the way for a future
In Table III. some grouping types are shown with some where product recommendations resonate more
groupings, these grouping types will define how accurately authentically with individual preferences. As the pursuit of
the recommendation is working. innovation continues, the potential for enhancing online
shopping experiences remains substantial, offering exciting
Table 3: Working of Grouping Metric in System avenues for future exploration and development in this
GROUPING TYPE GROUPING VALUE domain.
Brand Casio
. Titan REFERENCES
. FastTrack
[1]. Charles Ntumba, Samuel Aguayo, Kamau Maina
Age Group Kids-unisex
"Revolutionizing Retail: A Mini Review of E-
. Adults-men commerce Evolution" (2023).
. Kids-girls [2]. Devi Sunuwar; Monika Singh " Comparative
Analysis of Relational and Graph Databases for Data
D. History of Traversal Provenance: Performance, Queries, and Security
In Table IV each user's search history will be Considerations " (2023).
maintained and also what grouping values they landed to. [3]. Pablo Sánchez , Alejandro Bellogín " Point-of-
This will help us to give specific recommendations about Interest Recommender Systems Based on Location-
user's choice. Based Social Networks: A Survey from an
Experimental Perspective" (2022).
Table 4: Working of History Metric in System [4]. Breese, John S., David Heckerman, and Carl Kadie.
SER ID PRODUCT GROUPING "Empirical analysis of predictive algorithms for
VISITED VALUES collaborative filtering." arXiv preprint
U1 #U764 Casio arXiv:1301.7363 (2013).
. Adults-men [5]. Sarwar, B. M., Karypis, G., Konstan, J., & Riedl, J.
. Black (2002). Recommender systems for large-scale e-
. Accessories commerce: Scalable neighborhood formation using
U1 #U892 Kids-unisex clustering. In Paper presented at the Proceedings of
. Titan the fifth international conference on com- puter and
. FastTrack information technolog.
. Accessories [6]. D. Clark, IBM and Twitter forge partnership in data
analytics, Wall Street J. Retrieved April 15, 2016 from
http://www.marketwatch.com/story/ibm-andtwitter-
forge- partnership-on-data-analytics-2014-10-29.

IJISRT24APR1249 www.ijisrt.com 2581

Volume 9, Issue 4, April – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24APR1249

[7]. H.J. Watson, O. Marjanovic, Big data: the fourth data

management generation, Bus. Intelligence J. 18(3)
(2013) 4–8 (Chicago)
[8]. E. Dumbill, Making sense of big data, Big Data 1 (1)
(2013)1–2.
[9]. M. Anshari, Y. Alas, N. Yunus, N.I. Sabtu, M.H.
Hamid, Social customer relationship management
and student empowerment in online learning
systems, Int. J. Electronic Customer Relat. Manage.
9(2/3) (2015) 104–121.
[10]. H. Hinchcliffe, The state of Web 2.0, 2006.
Retrieved 12thMay,2012;from
http://web2.socialcomputingmagazine.com/the_state_
of_web_2 0.htm
[11]. P. Greenberg, CRM at the Speed of Light: Social
CRM 2.0 Strategies, Tools, and Techniques for
Engaging yOur Customers, fourth ed., McGraw-Hill
Osborne Media, 2009.
[12]. De Gemmis, M., Lops, P., Semeraro, G., & Basile, P.
(2008). Integrating tags in a semantic content-based
recommender. In Proceedings of the 2008 ACM
conference on Recommender systems (pp. 163–170).
ACM.
[13]. Gedikli, F., & Jannach, D. (2013). Improving
recommendation accuracy based on item-specific tag
preferences. ACM Transactions on Intelligent
Systems and Technology (TIST), 4, 11.
[14]. Liu, J., Wang, W., Chen, Z., Du, X., & Qi, Q. (2012).
A
[15]. novel user-based collaborative filtering method by
inferring tag ratings. ACM SIGAPP Applied
Computing Review, 12 , 48–57
[16]. Koren, Y., Bell, R., & Volinsky, C. (2009). Matrix
factorization techniques for recom- mender systems.
Computer, 42(8), 30–37.
[17]. Enrich, M., Braunhofer, M., & Ricci, F. (2013). Cold-
start management with cross-domain collaborative
filtering and tags. In International Conference on
Electronic Commerce and Web Technologies (pp.
101–112). Springer.
[18]. Bao, T., Ge, Y., Chen, E., Xiong, H., & Tian, J.
(2012).
[19]. Collaborative filtering with user ratings and tags. In
Proceedings of the 1st International Workshop on
Context Discovery and Data Mining (p. 1). ACM.
[20]. Wang, Z., & He, L. (2016). User identification for
enhancing ip-tv recommendation. Knowledge-Based
Systems, 98, 68–75.

IJISRT24APR1249 www.ijisrt.com 2582