The Progress of Shopping Recommendation System Based on

Machine Learning Algorithms

Xiyue Xiong

Data Science and Big Data Technology, University of Reading, Reading, U.K.

Keywords: Machine Learning, Big Data, Random Forest, Recommendation System, Deep Learning Model.

Abstract: The shopping recommendation systems play a crucial role in enhancing the online shopping experience by

providing personalized suggestions to users, thereby significantly increasing customer satisfaction and loyalty.

This paper discusses the application of machine learning in the context of the commercial shopping field,

especially for the shopping recommendation system. The algorithms discussed include random forest,

collaborative filtering, deep learning models, deep convolutional networks and other algorithms to elaborate

on its application, describing the important role played by machine learning technology in the commercial

field. It describes the important role that machine learning technology plays in the commercial field, and

makes an outlook assessment of the future development of this technology. Data science and related computer

technologies are used frequently and with great importance in the production of recommendation systems

when shopping. Different countries are also looking at this as a hot technology and are actively exploring and

developing algorithms for it.

1 INTRODUCTION

The rise of e-commerce has changed people's

consumption patterns, an increasing number of

individuals switch from offline shopping to online

shopping, placing orders and making payments

through the Internet. Today's Internet users have

access to a wealth of information because of the

growth and popularisation of the Internet industry,

their demands for information are met. However,

users have to face the excessive amount of

information and they cannot get the information that

is really useful for them, the efficiency of the use of

information has been reduced. With the increase in

the variety and number of products, users will browse

a lot of unwanted products. How to help users find the

products they need quickly and enhance user

viscosity is the focus of many online shopping malls

in the face of the globalisation of technology and

networking. In this context, personal

recommendation system, which recommends

information, products, etc. based on the user's

information needs, interests, etc. A smart

recommendation system not only provides

https://orcid.org/0009-0009-5469-1197

personalised services to users, but also builds a close

relationship with them so that they become reliable

on the recommendations.

Big data recommendation algorithms first

originated in Europe as a systematic optimization of

hadoop algorithms. Joldzic worked on large datasets

on Hadoop clusters Pessemier started working on

recommendation algorithms on Hadoop systems and

Mapreduce frameworks in 2011 distributed

processing and recommendation modeling research.

Currently, research on big data recommender system

algorithms in various countries focuses on

collaborative filtering algorithms, high-performance

computing recommendations, hybrid

recommendations, and algorithmic combinations.

However, in terms of research areas, the focus varies

from country to country. In China, researchers focus

on further optimization of the algorithm itself. Such

as relationship trustworthiness, user nearest neighbor,

matrix analysis, BP neural network, etc. Other

countries have focused more on the use of

recommender systems. There are a wide range of

applications, such as healthcare, online education,

social networking services, and recommender

systems are particularly popular for e-shopping.

110

Xiong, X.

The Progress of Shopping Recommendation System Based on Machine Learning Algorithms.

DOI: 10.5220/0012910800004508

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

In Proceedings of the 1st International Conference on Engineering Management, Information Technology and Intelligence (EMITI 2024), pages 110-114

ISBN: 978-989-758-713-9

This paper is dedicated to the application of

recommender systems in the field of shopping in a

globalized environment, in which in the first part, this

article introduces the research background of big data

analytics technology and its specific application in the

shopping field, the current research status of

recommender system algorithms in various countries,

and then in the second part, it describes the

introduction of various algorithms, such as machine

learning, and the way they are applied in shopping

recommender systems.

2 METHOD

Machine learning encompasses myriad applications,

ranging from high-performance implementations

such as artificial neural networks, random forests, to

support vector machines. To further elevate the

efficacy of these models, advanced deep learning

methodologies like deep neural networks, residual

neural networks, and deep forests have been

concurrently developed and implemented.

2.1 Machine Learning

The realm of machine learning revolves around the

exploration of ways computers can imitate or

replicate human learning behaviors, enabling them to

absorb new knowledge or skills, in addition to

reorganizing existing knowledge structures, thereby

facilitating improved performance over time. The

traditional applicability of machine learning

algorithms to small datasets is challenged by the sheer

volume and the intricate characteristics of big data.

Thus, machine learning algorithms hold immense

scholarly significance, being of immense relevance in

both the academic and industrial spheres.

Consequently, the investigation of machine learning

algorithms within the ambit of big data has surged in

popularity in both scholastic and commercial

domains.

2.1.1 Random Forest

Random Forest (RF), a central pillar of statistical

learning theory, utilizes the bootstrap resampling

method that extracts multiple specimens from the

original data sample. Subsequently, decision tree

modeling is employed on each bootstrap sample. The

predictions derived from various decision trees are

aggregated to generate a final prognosis through a

voting mechanism. In the context of the independent

variable X, each decision tree classification model

nominates the optimal classification outcome.

Thereby, the RF model comprises a composite

classification model, assembled from a multitude of

decision tree classification models. In the arena of

economic management, RF finds substantial utility,

particularly in forecasting potential customer

attrition. For instance, when employed in the domain

of customer relationship management, RF

demonstrated superior efficacy (Lariviere, 2005).

Further research suggested that Weighted Random

Forests eclipse standard RF in terms of predictive

capabilities, especially in the context of AUC metrics

(Burez, 2009). Beyond customer churn predictions,

RF has been deployed successfully in customer

loyalty forecasting. For example, Buckinx et al.

advocated the integration of customer loyalty

predictive values within a customer transaction

database, and evaluated the predictive capacities of

multiple linear regression, RF, and ANN (Buckinx,

2007).

2.1.2 Collaborative Filtering

The concept of "collaborative filtering" found its

origin in the mid-1990s when Goldberg and his

associates coined the term during the development of

their recommendation system, Tapestry (David,

1992). Since then, this method has been the subject of

extensive research and wide-ranging applications.

Collaborative filtering is founded on the basic

assumption that users A and B, who exhibit similar

historical annotation patterns or behavioral habits

(such as purchasing habits, reading preferences, film

choices, etc.), are likely to share similar interests on

other items. Generally, collaborative filtering

techniques employ a database to record users'

historical annotations, which is then leveraged to

predict user interests and offer personalized

recommendations. Amazon, a pioneering online

bookstore with no physical storefront, exemplifies

excellent use of collaborative filtering. It offers an

extensive database and a sophisticated search system

that allows users to conveniently look up book

information online. Users can add chosen books to a

virtual shopping basket, review their selected

products, choose their preferred service, and place

their order, thereby receiving their purchases at home

within a few days. Amazon's advanced personalized

recommendation function stands out, enabling it to

suggest books attuned to the diverse interests and

preferences of its users. This recommendation

software examines the books purchased by a reader,

as well as their evaluation of other books, to suggest

new books that the reader might like. Amazon's

The Progress of Shopping Recommendation System Based on Machine Learning Algorithms

111

system can analyze past purchases to make tailored

suggestions, saving customer information to allow for

more personalized suggestions during future visits.

Furthermore, Amazon's stellar post-sales service adds

to its appeal. Within 30 days of purchase, customers

can return their purchases in good condition to

Amazon for a full refund of the original price. Upon

return visits to the site, Amazon greets customers by

name, enhancing their online shopping experience.

2.2 Deep Learning

Deep learning, within the realm of signal processing,

has wide-ranging applicability, extending beyond

sounds, images, and videos to encompass text,

language, and the transmission of semantically

enriched information that humans can interpret. In the

domain of voice recognition, traditional Multilayer

Perceptrons (MLPs) have been utilised for an

extended period. However, standalone MLPs

significantly underperform compared to systems that

employ the Gaussian Mixture Model-Hidden Markov

Model (GMM-HMM). A breakthrough has been

observed in the challenge of large vocabulary

continuous speech recognition (LVC-SR) with the

innovative application of deep learning techniques.

This evolutionary leap is attributable to the

amalgamation of Hidden Markov Models (HMMs)

leveraging sequence modeling expertise, and Deep

Belief Networks (DBNs) offering robust

discriminative capabilities (Dahl,2011). Further

advancements were signaled by the efficient binary

coding of speech spectrograms using a deep auto-

encoder (Deng, 2010). This development further

emphasizes the extensive possibilities of deep

learning in signal processing.

2.2.1 CNN

A Convolutional Neural Network (CNN) is a

specialized artificial neural network designed to

process two-dimensional input data. Within a CNN,

each layer comprises numerous two-dimensional

planes, with each plane consisting of multiple

independent neurons. Neurons in two adjacent layers

are interconnected, while there are no connections

among neurons within the same layer. By employing

fewer network linkages and weight parameters

compared to fully connected networks of comparable

dimensions, CNNs effectively reduce the learning

complexity of network models, rendering them

simpler to train. A significant application of CNNs

can be seen in YOLO (You Only Look Once), which

approaches object detection as a regression problem

(Redmon, 2015). The YOLO model uses a CNN

structure with 24 convolutional layers and two fully

connected layers. It inputs the entire image and

divides it into 7*7 grids, predicting encapsulating

boxes and their class probabilities via CNN. YOLO

has the advantage of minimal background error and

high detection speed, capable of processing 45

images per second on a Titan GPU. Another

intriguing application of CNNs is in the Faster R-

CNN (Ren, 2016). It combines the Fast R-CNN

(Girshick, 2015) with a region proposal network,

sharing a convolutional layer feature network,

thereby improving object detection speed and

accuracy. Furthermore, CNNs have been employed

for image fusion (Li, 2016), short text clustering (Xu,

2015) and in numerous other domains.

3 DISCUSSION

3.1 Feature Extraction

While techniques for automatic feature extraction for

text information are fairly advanced, the extensive

amount of information available on the internet is

shared in multimedia formats. Recommending solely

textual information falls short in satisfying user needs.

Progress in multimedia information recommendation

research has been gradual, primarily due to the

constraints of automatic feature extraction

technology for such information. Presently,

multimedia information recommendations are often

generated based on manual annotations by users, a

process that also faces the issue of overfitting when

recommending text-based information. Hybrid

recommendation methods can bolster the diversity of

content-based recommendations, addressing these

challenges.

3.2 Scalability Problems

When user numbers for a commercial recommender

system escalate into millions or even tens of millions,

scalability problems pose a significant challenge for

the recommendation algorithm. Many online

recommendation platforms necessitate swift

provision of recommendation results to users, which

imposes stringent expectations for the timeliness of

recommendations. Regrettably, most extant

recommendation algorithms lack scalability. Some

relief from scalability challenges can be achieved by

employing strategies such as dimensionality

reduction, clustering, and classification. For instance,

dimensionality reduction techniques like Singular

EMITI 2024 - International Conference on Engineering Management, Information Technology and Intelligence

112

Value Decomposition (SVD) can compress matrices

and yield improved recommendation results.

However, matrix decomposition is time-consuming.

Algorithms like the nearest-neighbor-based KNN

(Deshpande, 2004), which considers only the closest

neighbors exhibiting the highest similarity to a target

user, can somewhat decrease the time overhead of the

recommendation process. Users are categorized

based on their preferences, allowing for

recommendations to be based predominantly on users

sharing similar preferences with the target user.

Model-based collaborative filtering (CF) algorithms,

like clustering collaborative filtering algorithms

categorize users by their preferences (Conno, 1999;

Sarwar, 2002), factoring in only users from the same

category as the target user during the

recommendation process. These methods can

navigate scalability issues, bolstering the

performance and utility of recommender systems.

3.3 Data Sparsity Problem

Data sparsity poses a major challenge for effective

recommender systems. Collaborative filtering

recommendation algorithms rely heavily on user-item

rating matrices, but these matrices are typically sparse,

leading to potential inaccuracies in generated

recommendations. This problem is compounded

when new users enter the system. Due to a lack of

rating data for such users, the generation of

personalized recommendations becomes challenging,

a predicament known as the cold start problem.

Dimensionality reduction techniques offer a common

solution to this issue (Bilus, 1998). These methods

compress matrices - for instance, through the use of

singular value decomposition to eliminate

unimportant or noisy users and items, thereby

reducing the dimensionality of user-item rating

matrices. Latent indexing techniques are also used to

project users' dimensions into a lower latitude space,

facilitating the calculation of user similarities.

Another viable solution to mitigate data sparsity is to

generate recommendations based on social data. This

approach leverages existing data, thereby enabling

fuller and more accurate user profiles for

recommendation generation (Bilus, 1998; Canny J,

2022)

3.4 Other Issues

Besides the central challenges, recommender systems

grapple with additional issues such as privacy

concerns (Feng, 2006). Many users are hesitant to

share their ratings or historical browsing data. The

resolution to this issue lies in enhancing the

trustworthiness of recommender systems and

designing systems that safeguard users' private data.

Furthermore, when introducing recommendations to

users, it's crucial to provide a justification for the

recommended product (Herlocker, 2000). For

instance, Amazon furnishes users with reasons like

"recommended for you because you browsed/bought

this". However, a significant number of present

recommendation algorithms fall short in providing

satisfactory rationales for their recommendations.

4 CONCLUSIONS

The study reviews pertinent algorithms, scrutinizes

their strengths and weaknesses, delves deeper into

commonly employed strategies for scoring and

similarity computations in recommendation

algorithms, and encapsulates the evaluation criteria

and methodologies typically used for assessing

recommendation algorithms. Though

recommendation technology has seen substantial

advancements over the last two decades, it still falls

short of an all-encompassing resolution to the core

issue. As the scope of application domains expands,

recommender systems will grapple with fresh

demands and challenges. Investigations in intelligent

information processing realms, such as information

retrieval, remain focused on devising recommender

solutions for the aforementioned problems. This is

because the evolution of recommender systems and

the challenges they face remain inherently

intertwined.

REFERENCES

Bilus D, Pazzani M, 1998, Learning collaborative

information filters. Process of International Conference

on Machine Learning, 46-54.

Buckinx W, Vverstraeten G, Poel D V D, 2007, Predicting

customer loyalty using the internal transactional

database. Expert Systems with Applications, 32(1).

Burez J, Poel DV D, 2009, Handling class imbalance in

customer churn prediction. Expert Systems with

Applications, 36(3).

Canny J, 2022, Collaborative filtering with privacy via

factor analysis. Process of the 25th International ACM

SIGIR Conference on Research and Development in

Information Retrieval, 238-245.

Connor M O, Herlocker J L, 1999, Clustering items for

collaborative filtering. Process of the ACM SIGIR

Workshop on Recommender Systems.

The Progress of Shopping Recommendation System Based on Machine Learning Algorithms

113

Dahl G E, et al., 2011, Large vocabulary continuous speech

recognition with context-dependent DBN-HMMs. Proc

of IEEE International Conference on Acoustics, Speech

and Signal Processing, 4688-4691.

Deng L, et al., 2010, Binary coding of speech spectrograms

using a deep auto-encoder. Proc of the 11th Annual

Conference of International Speech Communication

Association, 1692-1695.

Deshpande M, Karypis G, 2004, Item-based top-N

recommendation algorithms. ACM Transaction on

Information Systems, 22(1): 143-177.

Feng Z, Huiyou C, 2006, A study on privacy preserving

collaborative filtering recommendation based on

distributed data. Journal of Computing, 29(8): 1487-

1495.

Girshick R, 2015, Fast R-CNN. Proceedings of the 2015

IEEE International Conference on Computer Vision.

Santiago, Chile.

Goldberg D, Nichols D, Oki BM, Terry D, 1992, Using

collaborative filtering to weave an information tapestry.

Communications of the ACM, 12.

Herlocker JL, Konstan JA, Riedl J, 2000, Explaining

collaborative filtering recommendations. Process of the

ACM Conference on Computer Supported Cooperative

Work.

Lariviere B, Poel DVD, 2005, Predicting customer

retention and profitability by using random forests and

regression forests techniques. Expert Systems with

Applications, 29(2).

Li H, Liu F, Yang S Y, et al., 2016, Remote sensing image

fusion based on deep support value learning networks.

Computers, 39(8): 1583-1596 (in Chinese).

Redmon J, Divval S, Girshick R, et al., 2015, You only look

once: Unified, real-time object detection.

arXiv:1506.02640v5.

Ren S Q, et al., 2016, Faster R-CNN: Towards real-time

object detection with region proposal networks. IEEE

Transactions on Pattern Analysis and Machine

Intelligence, doi:10.1109/TPAMI.2016.2577031.

Sarwar B M, Karypis G, Konstan J A, et al., 2002,

Recommender systems for large-scale e-commerce:

Scalable neighborhood formation using clustering.

Process of the 5th International Conference on

Computer and Information Technology.

Xu J M, et al., 2015, Short text clustering via convolutional

neural networks. Proceedings of the NAACL-HLT

2015. Denver, USA, 62-69.

EMITI 2024 - International Conference on Engineering Management, Information Technology and Intelligence

114