Exploit Multi Layer Deep Learning and Latent Factor to Handle Sparse

Data for E-commerce Recommender System

Hanaﬁ

, Nanna Suryana Herman

and Abdul Samad Hasan Basari

Departement of Computer Science University of Amikom Yogyakarta Yogyakarta, Indonesia

Faculty of Information and Communication Technology University Teknikal Malaysia Malaka Malaka, Malaysia

Keywords:

E-commerce, Recommender system, Deep learning, LSTM, Sparse data, Collaborative ﬁltering.

Abstract:

E-commerce service have become popular way to shopping in recent decade. E-commerce machine requires

a method to provide ﬁt product information to customer called recommender system. The most of popular

recommender system adopted for many large e-commerce companies named Collaborative ﬁltering (CF) due

to obtain relevant, ﬁt and essential product information. Even though CF owned some beneﬁt, it has shortcom-

ing inaccurate recommendation when face minimum rating also popular named sparse data problem. Many

researches have been conducted to proposed model how to generate rating prediction aim to handle sparse rat-

ing effectively. Most of them exploit latent factor model or matrix factorization (MF) to handle this problem,

unfortunately, this problem fails to handle the problem when faced serious sparse data. Aims to improve the

serious problem on above, several researchers involve auxiliary information in the form of product document

or user demographic information respectively. Several researchers implemented Convolutional Neural Net-

work (CNN) to extract product document review incorporate MF that responsible to produce rating prediction,

another model exploited Stack Denoising Auto Encoder (SDAE) model as user demographic information ex-

traction incorporate with MF. In this research, considered implementing dual information representation using

deep learning model based on SDAE and Long Shorts Term Memory (LSTM) as product review document

representation combined into PMF to generate rating prediction. According to experiment report, the pro-

posed model called SLP (SDAE+LSTM+PMF) successful to obtained effectiveness rating prediction based

on RMSE evaluation metrices over some current model based on traditional PMF more than 15% in average

and superior over CNN more than 0.9% in average.

1 INTRODUCTION

Recommender system is one of the most important

tools to build success e-commerce business company.

Successful applied recommender system, it would in-

ﬂuent the selling target achievement in online trans-

action. Many large world companies have been im-

plementing recommender system to increase satisﬁed

service to their company to make customer enjoyable

looking for the product. It is a essential equipment

to promote sales and services for many online web-

sites and mobile applications. For instances, 80 per-

cent of movies watched on Netﬂix came from rec-

ommendations (Gomez-Uribe and Hunt, 2015), 60

percent of video clicks came from home page rec-

ommendation in YouTube (Davidson et al., 2010).

According to (Schafer et al., 2001) found that sales

agents with recommendations by the NetPerceptions

system achieved 60% higher average cross-sell value

and 50% higher cross-sell success rate than agents us-

ing traditional cross-sell techniques based on experi-

ments conducted at a U.K.-based retail and business

group.

Based on algorithm approach (Hanaﬁ et al., 2018),

e-commerce recommender system divided into 4

types as follow: 1). Content based: the method to

generate recommendation according to product clas-

siﬁcation approach. In the other hand, it is tend-

ing information retrieval to generate product recom-

mendation 2). Knowledge based: this approach is

to develops for speciﬁc necessary recommendation,

the speciﬁc character is to provide product informa-

tion rarely needed for individuals purpose for exam-

ple house, loan, insurance, car. 3). Demographic

based: product recommendation in which established

to provide product recommendation according to de-

mographic information. 4). Collaborative Filter-

ing (CF): the mechanism to produce recommendation

Hanaﬁ, . and Basari, A.

Exploit Multi Layer Deep Learning and Latent Factor to Handle Sparse Data for E-commerce Recommender System.

DOI: 10.5220/0009910603430351

In Proceedings of the International Conferences on Information System and Technology (CONRIST 2019), pages 343-351

ISBN: 978-989-758-453-4

343

based on user attitude in the past such as rating, prod-

uct review, comment, testimony, purchasing and etc.

CF is the most successful technique that implemented

in many large e-commerce company due to CF have

ability to provide recommendation with special char-

acter as follow; provide product ﬁt, serve relevant in-

formation, accurate, serendipity (Ricci et al., 2015).

In common use, CF adopted rating as explicit feed-

back as basic compute to calculate similarity user for

product uses rating matrix to generate product recom-

mendation. The big challenge in collaborative ﬁlter-

ing is just slight of user population who gave rating

for product, totally about only less than 1 percent.

The problem popular called sparse data also in ex-

treme condition sparse data famous called cold start.

When cold start happens, there is no recommendation

possible to generated by recommendation system.

CF having several advantages over another ap-

proach such Content base, Demographic base,

Knowledge base. However, CF have essential limita-

tion caused lack in rating collection from customers.

CF rely on rating as basic calculation to generate

product recommendation. Several attempts have been

made to reduce sparse data, so they hope the product

recommendations method having more accurate. The

use of auxiliary information has proven to improve

the accuracy of a product recommendation. Some of

the auxiliary information that have been explored to

handle sparse data have been done for instance audio

features in music recommendations (Van den Oord

et al., 2013) (Wang and Wang, 2014), color features

on online fashion shop (Jaradat, 2017), documents

recommendation in online news (Park et al., 2017).

With the ﬁnal objective to beat the issue of in-

formation ﬂooding, utilizations of recommendation

system have pulled in a lot of consideration in last

decade. Numerous business companies have included

recommender system into their e-commerce intelli-

gent system, for example, Amazon1, JD2, Taobao3,

and, etc. One viable technique for recommendation

is to estimate new rating of various users and items.

Presently, the most famous strategies in this ﬁeld are

Content-based Filtering (CBF) and Collaborative Fil-

tering (CF). CBF utilizes the setting of users or items

to estimate new rating. For instance, we can create

user inclinations as indicated by their age, sexual ori-

entation, or diagram of their colleague, and etc. Like-

wise, genres and items review can be exploited to gen-

erate product preference for various users. On the op-

posite side, CF utilizes the rating of users for items in

their review to estimate new ratings. For example, a

table of N users {u

, u

, ...u

} and table of

M items {v

, v

, ...v

} . Concurrently, a ta-

ble of items v

, given rated by users u

. These ratings

can either be explicit feedback on rating star 1-5, or

implicit feedback on a scale of 0-1. Our objective is

to estimate new feedback from users without records.

Moreover, CF often more powerful over CBF due CF

(F. Ricci and Saphira, 2011). Fig. 1 is example of

movie review from users that very popular to inte-

grated into latent factor model in the term of matrix

factorization.

Figure 1: example review product for movie

One of the most outstanding methodologies in CF

are Probabilistic Matrix Factorization (PMF) (Mnih

and Salakhutdinov, 2008) and Singular Value Decom-

position (SVD) (Sarwar et al., 2001). However, they

are usually against with sparse data (extreme sparse in

cold start issue), so they stall to squeeze effective de-

scription from users and items. To address the trouble

related with this issue, a functioning line of research

since the previous decade and several of method have

been proposed (Zhou et al., 2011) (Yi et al., 2016)

(Trevisiol et al., 2014). All of these methods consider

utilizing the side information of users or items into CF

to produce more effective characteristic. In addition,

the most normally utilized side information text doc-

uments, therefore several methods use text document

approach, for example, Latent Dirichlet Allocation

and Collaborative Topic Modelling have been put for-

ward by (Ling et al., 2014), (Wang and Blei, 2011)hile

earlier research has inspected text document as a bag-

of-words, it might be desirable over ponder the ef-

fect of sequences of words in text documents; con-

sequently, past techniques are constrained to several

expand.

Figure 2: Rating given by users.

According ﬁg. 2, majority problem of sparse data

problem caused minimum rating.

CONRIST 2019 - International Conferences on Information System and Technology

344

2 PREVIOUS WORK

DDeep learning is one of the machine learning ap-

proaches that used to solve data sparse problems

and overcome some of the weaknesses in several ap-

proaches described above. Deep learning is a deriva-

tive of a resurgent neural network because it can pro-

duces outperform result in image processing, natural

language processing. Some experts attempt to em-

power deep learning in dealing with problems that ex-

ist in the recommender system. Several study was use

deep learning to extract content feature aims to han-

dle sparse data in rating product, such as (Van den

Oord et al., 2013) enhance Deep Convolutional Neu-

ral Network to create audio music classiﬁer aim to de-

velop auto music recommendation, due no data prod-

uct record, the recommendation difﬁcult to generate.

According (HANAFI et al., 2018) where author

conducting approach to eliminate sparse data by com-

bining between multi-layer neural network and non-

negative matrix factorization, even the result of hy-

bridization could be work, however the results are less

accurate than some studies that use similar technique.

space

Figure 3: comparison of state-of-the-art approach.

Source of Figure 3 : (Mnih and Salakhutdinov,

2008), (Wang and Blei, 2011), (Zhang et al., 2006),

(Wang et al., 2015b), (Sarwar et al., 2002), (Park

et al., 2017), (Liu et al., 2017).

3 OUR CONTRIBUTION

Since last decade deep learning have become trend-

ing to solving in several computer science research

ﬁeld because of outstanding improvement. Feature

extraction of deep learning has solid characteristic,

the increase in the number of studies has adopted

deep learning technology by implementing side in-

formation to produce effective description. For exam-

ple, Deep Recurrent Neural Network (RNN) for News

Recommendation (Park et al., 2017), another Author

proposed a model to dig user rating latent factor us-

ing Stacked Denoising Autoencoder (SDAE) (Wang

et al., 2015a) (Wang et al., 2015b) and Convolutional

Neural Network (CNN) (Wang et al., 2015a) a.k.a.

Exploit Multi Layer Deep Learning and Latent Factor to Handle Sparse Data for E-commerce Recommender System

345

ConvMF, which difference angle the inquiry of words

in text documents. In any case, ConvMF just thinks

about item side information (e.g., text documents, re-

view, synopsis, abstract, etc.), so users’ latent factors

still have no effective description. Another approach

by Author (Kim et al., 2016) shows that SDAE is

proﬁcient at extricating remarkable feature in users’

latent factor without involving text documents, this

causes items latent factor stay with equal classical

methodologies. Hence, according to handle the prob-

lem on above, we incorporate SDAE-NN and CDNN

into a probabilistic model to develop more effectively

extract users’ latent factor and items latent factor.

1. We propose hybrid model involve users latent

factor layer by using SDAE-NN and items la-

tent factor by using DC-NN, both of them incor-

porating by probabilistic model. Our proposed

model called MultiLayer Deep Learning (MLDL)

recommender system. To the best our insight,

MLDL is the pioneer model to incorporating two

deep learning layers (SDAE-NN and DC-NN)

into probabilistic point of view.

2. We widely show that MLDL is a mixing of a sev-

eral best performance techniques but with a more

effective representation.

3. We establish unique strategy and conduct exami-

nations which demonstrate that MLDL successful

to eliminate CF sparse data issue.

4 PROBLEM DEFINITION

Similar with some existing state of the art, this ex-

periment adopts document of product review to inte-

grated into probabilistic matrix factorization to pro-

duce rating prediction (PMF). Following to famous

pre-processing procedure, we have n users, m items,

and an extreme sparse in rating matrix R ∈ R

nxm

. Ev-

ery input R

i j

of R corresponds to user’s i rating on

item j. Involvement document product review explain

in below, the auxiliary information of users and items

are denoted by X ∈ R

nxe

and Y ∈ R

mx f

, respectively.

Let u

, v

∈ R be user i latent factor vector and item j

latent factor respectively, where k is the dimensional-

ity of latent space. In this process, the corresponding

matrix form of latent factor for users and items are U

= u[1:n] and V = v[1:m], separately. Given the sparse

rating matrix R and the side information matrix X and

also Y, our objective is to learn effective users latent

factor U and Item latent factor V, and after that to es-

timate the missing rating in R.

5 OUR APPROACH

In this section, we consider explaining three essen-

tial method to develop SLP (SDAE-LSTM-PMF) to

produce rating prediction to handle sparse data in rec-

ommender system based on collaborative ﬁltering.

5.1 A. Probabilistic Matrix

Factorization (PMF)

PMF is traditional latent factor model using matrix

factorization to produce rating prediction. This model

considers Gaussian Normal Distribution to create user

and item latent factor representation. First model pro-

posed by Salakhutdinov (Mnih and Salakhutdinov,

2008). Unfortunately, majority collaborative ﬁltering

based on traditional latent factor whether PMF, SVD,

SVD involve temporal effect, Non-Negative Matrix

Factorization (NNMF) obtain inaccurate rating pre-

diction when faced with sparse data. Aim to handle

this problem, considering item or user side informa-

tion are needed. In this experiment, considered PMF

required to obtain rating prediction integrated with

SDAE and LSTM.

5.2 SDAE to Extract User Demographic

Information

SDAE is sub class of neural network model where it

is based on feed forward approach. This model very

popular to adopt in deep learning mechanism. Ba-

sic concept of this approach by follow the term of

autoencoder mechanism by using feedforward neural

network that it has an input layer, one hidden layer

and an output layer as follow ﬁgure 5. The output

layer has same number of neurons as the input layer

for the purpose of reconstructing its own inputs. This

makes an autoencoders a form of unsupervised learn-

ing, which means no labelled data are necessary just a

set of input data instead of input-output pairs. In SLP

model (SDAE-LSTM-PMF, SDAE consider extract-

ing user demographic information representation to

transform within 2D latent space. User demographic

information were used as input of SDAE for feature

learning representation. The result of SDAE output in

the term of 2D latent space would be integrated with

PMF. Before SDAE processes be conducted, user de-

mographic information would be pre-processing us-

ing vectorized to transform user information into vec-

tor space.

CONRIST 2019 - International Conferences on Information System and Technology

346

5.3 Long Shorts Term Memory (LSTM)

to Extract Product Review

Document

TItem document information has become popular in-

formation to support latent factor model to improve

performance in producing recommendation in last

decade. Majority of previous work adopted LDA

model to interpreted product review document. Some

of researcher using CNN to make more deeper under-

standing of product document interpretation. How-

ever, several of them fail to capture contextual un-

derstanding of product review document due to ig-

nore sequential word aspect to develop latent seman-

tic representation. Aims to improve the problem in

contextual understanding, this research considers to

adopted LSTM model to detect contextual understat-

ing of product review document.

LSTM method adopted to transform item docu-

ment product review. The detail explanation of our

model shows on Fig 3 below where item document

obtains from IMDB, then pre-processing process us-

ing NLTK (Natural Language Tool Kits) module con-

duct some process such as lemmatization, stop word

removal, remove punctuation and vectorized into one

hot encoding. After pre-processing stage success

to implemented, exploiting LSTM using TensorFlow

would be utilities in the next processes. The ﬁnal

process resulted 2D latent space in the form 50 di-

mensional. So, the document product review as prod-

uct representation would integrate with SDAE as user

information representation into latent factor model

based on PMF.

Figure 4: LSTM to extract item document information

5.3.1 Hybrid SDAE, LSTM and PMF

The success story of model based using latent fac-

tor based on matrix factorization popularized several

researchers and academicians in Netﬂix competition

event on early 2006. Majority of matrix factoriza-

tion success to improve memory-based model signif-

icantly. Began in this competition, model based us-

ing matrix factorization become favorited model to

increase performance of collaborative ﬁltering recom-

mendation.

There are several researches considered to inte-

grated deep learning model with matrix factorization

to increase performance in handling sparse data due to

deep learning has tremendous achievement for several

computer sciences ﬁeld such Audio, image process-

ing, sentiment analysis, text mining. A kind of deep

learning class is convolutional neural network, feed-

forward neural system that has effectively been em-

ployed to Computer Vision (Krizhevsky et al., 2012),

Natural Language Processing Kim (Kim, 2014) and

Audio Signal Processing (Piczak, 2015). Essen-

tially, there are a few examinations have utilized CNN

in recommender system, for example, content-based

music recommendation (Van den Oord et al., 2013),

fashion shop based colour feature (Jaradat, 2017),

which opposed with conventional method utilizing a

bag of words as description of the audio signal with

deep CNN, and Dynamic Convolutional Neural Net-

work (DCNN), which utilizes CNN to produce items’

latent factors however this model lack in users latent

factor due the users security issue. In other hand,

their effort just thinks one sided latent factor (i.e.,

item auxiliary information), which uses estimated rat-

ing not similar to SVD via gradient descent. Hence,

we utilize items’ latent factor developed by LSTM

and user’ latent factor by a family of SDAE. We have

prepared SDAE model which consideration imputing

missing value to increase effective representation of

user latent factor. In the previous work, considera-

tion integrated user and item information represen-

tation into matrix factorization has been made that

called PHDMF. This model exploit CNN to capture

document of product review understanding. Different

with PHDMF, our approach exploits LSTM to cap-

ture contextual understanding of product review doc-

ument. The detail of SLM model demonstrated on

Figure 5 below.

Exploit Multi Layer Deep Learning and Latent Factor to Handle Sparse Data for E-commerce Recommender System

347

space

Figure 5: Integrate SLM model include SDAE, LSTM and

PMF

Figure 5 on above show overview of the proba-

bilistic matrix factorization approach to support SLP,

which combine SDAE and LSTM model into PMF.

Based on a probabilistic point of view, the formula

given by:

p(R|U, V, σ

) =

∏

N(R

i j

, σ

)

i j

(1)

Where N(x|µ , σ

) is the probability density func-

tion of the Gaussian normal distribution with mean µ

and variance σ

. Thus, to ﬁnalize users’ latent factor

can be produce by reference equation on below.

= sdae(W

, X

) + ε

= N(0, σ

I) (2)

To calculate w

inW

we consider using zero-

mean spherical Gaussian prior, equation according on

below.

P(w

|σ

) =

∏

N(w

|0, σ

) (3)

Finally, the distribution over user latent factor

given by:

p(U|W

, X, σ

) =

∏

N(u

|sdae(W

, X

), σ

) (4)

Similarly, the user’s latent factor, an item latent

factor given by formulation structure:

= lstm(W, Y

) + ε

= N(0, ε

I) (5)

dcnn representation of the output of dcnn ap-

proach, also use zero-mean spherical Gaussian prior.

P(wσ

∏

N(N

, 0, σ

)) (6)

p(V |W, Y, σ

) =

∏

N(v

|dcnn(W, Y

), σ

I) (7)

5.4 Evaluation Measure Metric

RMSE is frequently used measure of the differences

between values (sample and population values) pre-

dicted by a model or an estimator and the values ac-

tually observed. Root Mean Squared Error (RMSE)

is might the most famous metric used in evaluating

accuracy of predicting rating. The system generates

predicted ratings ¯r

for test set t of user-item pairs (u,i)

for which the true rating r

are known. Typically, r

are known because they are hidden in an ofﬂine exper-

iment. The RMSE between the predicted and actual

rating is given by:

RMSE =

|t|

(u,i)∈t

(¯r

− r

)

(8)

6 EXPERIMENT SETTING

This research focus to develop a method to handle

user rating sparse data extremely in case collaborative

ﬁltering recommender system involve deep learning

machine to ﬁnd relationship latent factor between user

rating from a product. According our best knowledge,

in state of the art by use matrix factorization has fail-

ure to addressing user rating sparse data extremely.

So, the result of recommendation sometime getting

the mistake (Koren et al., 2009). This research is

categorical lab scale research involve public datasets.

We consider MovieLens dataset as convince datasets

were applied in so many study in recommender sys-

tem research ﬁeld, detail speciﬁcation refer to (Harper

and Konstan, 2015).

6.1 Device and Library Tools

Our experiment involves several tools include soft-

ware and hardware. There are several tools and soft-

ware that involve includes Python with several li-

braries such as tensor ﬂow for deep learning imple-

mentation and GeForce GTX 1001 for running con-

volutional neural networks supported by processors

that we use Xeon 2.4 Ghz.

CONRIST 2019 - International Conferences on Information System and Technology

348

space

Figure 6: Detail speciﬁcation device and library.

6.2 Datasets

In this research to understand the performance of SLP

model, we considered to implemented on three real

datasets obtain from Movielens and AIV. The detail

characteristic of datasets shows on Figure 7 below.

Figure 7: Characteristic of Datasets.

To show the effectiveness of our approach in terms

of rating prediction. Firstly, we applied real-world

datasets acquired from MovieLens (Harper and Kon-

stan, 2015) and Amazon 3. These datasets contain

of consumer’ explicit ratings for products on rating

scale of 1 to 5. Amazon dataset contains opinion for

products as item description documents. Because of

MovieLens does not include item description docu-

ments, we generate the documents use correspond-

ing items from IMDB server. Equal with (Liu et al.,

2017), we conduct to preprocessed description doc-

uments for all of them datasets as follows: 1) set-

ting the data with maximum length of raw documents

to 300, 2) removing stop words, 3) calculated tf-idf

score for each word, 4) removing corpus speciﬁc stop

words that have the document frequency higher than

0.5, 5) Choose top 8000 distinct words as a vocabu-

lary, 6) removing all non-vocabulary words from raw

documents. As a result, average numbers of words

per document are 97.09 on MovieLens-1M (ML-1M),

92.05 and Amazon Instant Video (AIV), We consider

to deleted items that have no their description doc-

uments in dataset table, and specially for the case

of Amazon dataset, we consider to removed user’s

data that have just only less than 3 ratings. The re-

sult, statistics of each data show that three datasets

have different characteristics on table 2. Finally, even

though some users have removed by preprocessing,

Amazon dataset is still extremely sparse compared to

the others.

7 RESULT

The result of experiment report shows on ﬁgure be-

low include some scenario training process where we

divided dataset into training and testing categories

with interval 10%. So, our scenario experiment in-

cluded 10/90, 20/80, 30/70, 40/60, 50/50, 60/40,

70/30, 80/20, 90/10. The complete experiment result

demonstrated ﬁgure below.

Figure 8: Evaluation metrices result ratio 10:90

Figure 9: Evaluation metrices result ratio 20:80

Figure 10: Evaluation metrices result ratio 30:70

Exploit Multi Layer Deep Learning and Latent Factor to Handle Sparse Data for E-commerce Recommender System

349

space

Figure 11: Evaluation metrices result ratio 40:60

Figure 12: Evaluation metrices result ratio 50:50

Figure 13: Evaluation metrices result ratio 60:40

Figure 14: Evaluation metrices result ratio 70:30

space

Figure 15: Evaluation metrices result ratio 80:20

Figure 16: Evaluation metrices result ratio 90:10

8 PRELIMINARY FINDING

Based on this experiment it can be shown our pro-

posed hybrid model called SLP success to produce

rating prediction using ninth scenario training and

testing process. According to the experiment report

on above, the best performs on ratio 90:10 composi-

tion training and testing compare to another compe-

tition. User information latent factor and item docu-

ment latent factor play important role to support PMF

in producing rating prediction. Our plan for future

experiment considers to implemented with several pa-

rameter and comparison with previous state of the art

to understand effectiveness rating prediction level.

CONRIST 2019 - International Conferences on Information System and Technology

350

REFERENCES

Davidson, J., Liebald, B., Liu, J., Nandy, P., Van Vleet, T.,

Gargi, U., Gupta, S., He, Y., Lambert, M., Livingston,

B., et al. (2010). The youtube video recommendation

system. In Proceedings of the fourth ACM conference

on Recommender systems, pages 293–296.

F. Ricci, L. R. and Saphira, B. (2011). Recommender sys-

tems handbook.

Gomez-Uribe, C. A. and Hunt, N. (2015). The netﬂix rec-

ommender system: Algorithms, business value, and

innovation. ACM Transactions on Management Infor-

mation Systems (TMIS), 6(4):1–19.

HANAFI, SURYANA, N., BASARI, A. S. B. H., et al.

(2018). Hybridization approach to eliminate sparse

data based on nonnegative matrix factorization & deep

learning. Journal of Theoretical & Applied Informa-

tion Technology, 96(14).

Hanaﬁ, SURYANA, N., BASARI, H., BIN, A. S., et al.

(2018). An understanding and approach solution for

cold start problem associated with recommender sys-

tem: A literature review. Journal of Theoretical &

Applied Information Technology, 96(9).

Harper, F. M. and Konstan, J. A. (2015). The movielens

datasets: History and context. Acm transactions on

interactive intelligent systems (tiis), 5(4):1–19.

Jaradat, S. (2017). Deep cross-domain fashion recommen-

dation. In Proceedings of the Eleventh ACM Confer-

ence on Recommender Systems, pages 407–410.

Jolly, P. (1982). Nickel catalyzed coupling of organic

halides and related reactions.

Kim, D., Park, C., Oh, J., Lee, S., and Yu, H. (2016). Con-

volutional matrix factorization for document context-

aware recommendation. In Proceedings of the 10th

ACM Conference on Recommender Systems, pages

233–240.

Kim, Y. (2014). Convolutional neural networks for sentence

classiﬁcation. arXiv preprint arXiv:1408.5882.

Koren, Y., Bell, R., and Volinsky, C. (2009). Matrix factor-

ization techniques for recommender systems. Com-

puter, 42(8):30–37.

Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012). Im-

agenet classiﬁcation with deep convolutional neural

networks. In Advances in neural information process-

ing systems, pages 1097–1105.

Ling, G., Lyu, M. R., and King, I. (2014). Ratings meet

reviews, a combined approach to recommend. In Pro-

ceedings of the 8th ACM Conference on Recommender

systems, pages 105–112.

Liu, J., Wang, D., and Ding, Y. (2017). Phd: a probabilistic

model of hybrid deep collaborative ﬁltering for rec-

ommender systems. In Asian Conference on machine

learning, pages 224–239.

Mnih, A. and Salakhutdinov, R. R. (2008). Probabilistic

matrix factorization. In Advances in neural informa-

tion processing systems, pages 1257–1264.

Park, K., Lee, J., and Choi, J. (2017). Deep neural net-

works for news recommendations. In Proceedings

of the 2017 ACM on Conference on Information and

Knowledge Management, pages 2255–2258.

Piczak, K. J. (2015). Environmental sound classiﬁcation

with convolutional neural networks. In 2015 IEEE

25th International Workshop on Machine Learning for

Signal Processing (MLSP), pages 1–6. IEEE.

Ricci, F., Rokach, L., and Shapira, B. (2015). Recom-

mender systems: introduction and challenges. In Rec-

ommender systems handbook, pages 1–34. Springer.

Sarwar, B., Karypis, G., Konstan, J., and Riedl, J. (2001).

Item-based collaborative ﬁltering recommendation al-

gorithms. In Proceedings of the 10th international

conference on World Wide Web, pages 285–295.

Sarwar, B., Karypis, G., Konstan, J., and Riedl, J. (2002).

Incremental singular value decomposition algorithms

for highly scalable recommender systems. In Fifth in-

ternational conference on computer and information

science, volume 1. Citeseer.

Schafer, J. B., Konstan, J. A., and Riedl, J. (2001). E-

commerce recommendation applications. Data min-

ing and knowledge discovery, 5(1-2):115–153.

Trevisiol, M., Aiello, L. M., Schifanella, R., and Jaimes,

A. (2014). Cold-start news recommendation with

domain-dependent browse graph. In Proceedings of

the 8th ACM Conference on Recommender systems,

pages 81–88.

Van den Oord, A., Dieleman, S., and Schrauwen, B.

(2013). Deep content-based music recommendation.

In Advances in neural information processing sys-

tems, pages 2643–2651.

Wang, C. and Blei, D. M. (2011). Collaborative topic

modeling for recommending scientiﬁc articles. In

Proceedings of the 17th ACM SIGKDD international

conference on Knowledge discovery and data mining,

pages 448–456.

Wang, H., Shi, X., and Yeung, D.-Y. (2015a). Relational

stacked denoising autoencoder for tag recommenda-

tion. In Twenty-ninth AAAI conference on artiﬁcial

intelligence.

Wang, H., Wang, N., and Yeung, D.-Y. (2015b). Collab-

orative deep learning for recommender systems. In

Proceedings of the 21th ACM SIGKDD international

conference on knowledge discovery and data mining,

pages 1235–1244.

Wang, X. and Wang, Y. (2014). Improving content-based

and hybrid music recommendation using deep learn-

ing. In Proceedings of the 22nd ACM international

conference on Multimedia, pages 627–636.

Yi, P., Yang, C., Zhou, X., and Li, C. (2016). A movie

cold-start recommendation method optimized simi-

larity measure. In 2016 16th International Sympo-

sium on Communications and Information Technolo-

gies (ISCIT), pages 231–234. IEEE.

Zhang, S., Wang, W., Ford, J., and Makedon, F. (2006).

Learning from incomplete ratings using non-negative

matrix factorization. In Proceedings of the 2006 SIAM

international conference on data mining, pages 549–

553. SIAM.

Zhou, K., Yang, S.-H., and Zha, H. (2011). Functional

matrix factorizations for cold-start recommendation.

In Proceedings of the 34th international ACM SIGIR

conference on Research and development in Informa-

tion Retrieval, pages 315–324.

Exploit Multi Layer Deep Learning and Latent Factor to Handle Sparse Data for E-commerce Recommender System

351