Recommender system using deep learning and matrix factorization
The recommender system is part of machine learning that responsible to provide product recommendation to consumers in e-commerce. This system has been adopted by almost every e-commerce company in the world including Amazon, Alibaba, iTunes, Google, and Netflix. Collaborative filtering (CF) is the m...
Saved in:
Main Author: | |
---|---|
Format: | Thesis |
Language: | English English |
Published: |
2021
|
Subjects: | |
Online Access: | http://eprints.utem.edu.my/id/eprint/26098/1/Recommender%20system%20using%20deep%20learning%20and%20matrix%20factorization.pdf http://eprints.utem.edu.my/id/eprint/26098/2/Recommender%20system%20using%20deep%20learning%20and%20matrix%20factorization.pdf |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
id |
my-utem-ep.26098 |
---|---|
record_format |
uketd_dc |
spelling |
my-utem-ep.260982022-11-15T12:49:47Z Recommender system using deep learning and matrix factorization 2021 Hanafi H Social Sciences (General) HF Commerce The recommender system is part of machine learning that responsible to provide product recommendation to consumers in e-commerce. This system has been adopted by almost every e-commerce company in the world including Amazon, Alibaba, iTunes, Google, and Netflix. Collaborative filtering (CF) is the most popular recommender system method. This method calculated the user behaviour similarities in the form of ratings. However, in reality, the number of ratings is difficult to obtain (0.3% - 4%). In 2006 when Netflix competition was held, many experts consider to using matrix factorization (latent factor) model, and this model has proved to be better performance over previous work using statistical method. However, the matrix factorization performs degrade when implemented on extreme sparse data. Adding information in the form of product documents using Latent Dirichlet Allocation (LDA) is one way to solve this problem. However, the LDA method also faces obstacles to capture the documents contextual understanding which mostly of work by using LDA as categorical Bag of Word (BOW). A significant effort to capture the contextual meaning in Natural Language Processing (NLP) application is by considering the subtle words and the words order. Another problem is that most researchers involve only a part of the information to aid the performance of the matrix factorization, namely the user information or product information. Aiming to deal with the contextual problem of the document, this research proposes the use of the Long Short-Term Memory (LTSM) and word embedding (WE) based on Global Vector for Word Representation (GLOVE) method. According to the testing and evaluation report, the LTSM and GLOVE which has proven to be successful to capture the contextual meaning of the document in qualitative tests. Aiming to deal with the matrix factorization’s decrease in performance, this research proposes to integrate LSTM and Probabilistic Matrix Factorization (PMF). According to the experiment report, LSTM-PMF superior over existing best perform using Convolutional Neural Network (CNN) and PMF achieve 1.4% on average. Aiming to deal with hybridization between user and item information representation, this study implements Stack Denoising Auto Encoder (SDAE) and LSTM into PMF. The experiment report shows that this model outperforms over previous work achieve 0.92% on average. In the future, aiming to improve the performance in capturing understanding of product document, the use of bidirectional word vector representation and another variant of deep learning needs to be considered. 2021 Thesis http://eprints.utem.edu.my/id/eprint/26098/ http://eprints.utem.edu.my/id/eprint/26098/1/Recommender%20system%20using%20deep%20learning%20and%20matrix%20factorization.pdf text en public http://eprints.utem.edu.my/id/eprint/26098/2/Recommender%20system%20using%20deep%20learning%20and%20matrix%20factorization.pdf text en validuser https://plh.utem.edu.my/cgi-bin/koha/opac-detail.pl?biblionumber=121352 phd doctoral Universiti Teknikal Malaysia Melaka Faculty of Information and Communication Technology Herman, Nanna Suryana |
institution |
Universiti Teknikal Malaysia Melaka |
collection |
UTeM Repository |
language |
English English |
advisor |
Herman, Nanna Suryana |
topic |
H Social Sciences (General) HF Commerce |
spellingShingle |
H Social Sciences (General) HF Commerce Hanafi Recommender system using deep learning and matrix factorization |
description |
The recommender system is part of machine learning that responsible to provide product recommendation to consumers in e-commerce. This system has been adopted by almost every e-commerce company in the world including Amazon, Alibaba, iTunes, Google, and Netflix. Collaborative filtering (CF) is the most popular recommender system method. This method calculated the user behaviour similarities in the form of ratings. However, in reality, the number of ratings is difficult to obtain (0.3% - 4%). In 2006 when Netflix competition was held, many experts consider to using matrix factorization (latent factor) model, and this model has proved to be better performance over previous work using statistical method. However, the matrix factorization performs degrade when implemented on extreme sparse data. Adding information in the form of product documents using Latent Dirichlet Allocation (LDA) is one way to solve this problem. However, the LDA method also faces obstacles to capture the documents contextual understanding which mostly of work by using LDA as categorical Bag of Word (BOW). A significant effort to capture the contextual meaning in Natural Language Processing (NLP) application is by considering the subtle words and the words order. Another problem is that most researchers involve only a part of the information to aid the performance of the matrix factorization, namely the user information or product information. Aiming to deal with the contextual problem of the document, this research proposes the use of the Long Short-Term Memory (LTSM) and word embedding (WE) based on Global Vector for Word Representation (GLOVE) method. According to the testing and evaluation report, the LTSM and GLOVE which has proven to be successful to capture the contextual meaning of the document in qualitative tests. Aiming to deal with the matrix factorization’s decrease in performance, this research proposes to integrate LSTM and Probabilistic Matrix Factorization (PMF). According to the experiment report, LSTM-PMF superior over existing best perform using Convolutional Neural Network (CNN) and PMF achieve 1.4% on average. Aiming to deal with hybridization between user and item information representation, this study implements Stack Denoising Auto Encoder (SDAE) and LSTM into PMF. The experiment report shows that this model outperforms over previous work achieve 0.92% on average. In the future, aiming to improve the performance in capturing understanding of product document, the use of bidirectional word vector representation and another variant of deep learning needs to be considered. |
format |
Thesis |
qualification_name |
Doctor of Philosophy (PhD.) |
qualification_level |
Doctorate |
author |
Hanafi |
author_facet |
Hanafi |
author_sort |
Hanafi |
title |
Recommender system using deep learning and matrix factorization |
title_short |
Recommender system using deep learning and matrix factorization |
title_full |
Recommender system using deep learning and matrix factorization |
title_fullStr |
Recommender system using deep learning and matrix factorization |
title_full_unstemmed |
Recommender system using deep learning and matrix factorization |
title_sort |
recommender system using deep learning and matrix factorization |
granting_institution |
Universiti Teknikal Malaysia Melaka |
granting_department |
Faculty of Information and Communication Technology |
publishDate |
2021 |
url |
http://eprints.utem.edu.my/id/eprint/26098/1/Recommender%20system%20using%20deep%20learning%20and%20matrix%20factorization.pdf http://eprints.utem.edu.my/id/eprint/26098/2/Recommender%20system%20using%20deep%20learning%20and%20matrix%20factorization.pdf |
_version_ |
1776103137946370048 |