Stochastic Modeling & Applications ISSN: 0972-3641

Vol. 26 No. 3 (January – June, Special Issue 2022 Part - 8) UGC CARE APPROVED JOURNAL

Received: 15th April 2022 Revised: 10th May 2022 Accepted: 8th June 2022

Assessing the Effectiveness of Hybridization Approach in Recommender Systems

Dr Y Padma1 ,Dr. S. Sai Kumar2 ,Dr.Shaik.Nagul3 ,Siyyadula V R Vara Prasad4
1
Assistant Professor ,Department of Information Technology ,PVP Siddhartha Institute of Technology
2
Assistant Professor ,Department of Information Technology ,PVP Siddhartha Institute of Technology
3
Associate Professor,Departmentof Cse,Lendi Institute of Engineering and Technology
4
Asistant Professor ,Lendi Institute of Engineering and Technology

ABSTRACT
A key element for effective E-commerce interactions between customers and sellers is the suggestion
framework. Sales outcomes are significantly impacted by quick and helpful contacts to discover the proper
product. Collective filtering, content-based, demographic, and programme model criteria make up the basis of
a technical method. The other methods on the list are thought to be inferior than collaborative filtering. It
benefits from necessity in terms of luck, freshness, and accuracy. The DLS Framework is a proposed system
that uses deep learning models for sentiment analysis for the DLS recommendation system. In terms of
performance across many domains and various real-time data volumes. The suggested models demonstrate
advantages and contrast the results with those of other models already in use.
Keywords: E-commerce, DL, DLS, Sentiment anlaysis.

1. INTRODUCTION
The best opinions and reviews on various product categories are found using recommender systems (RS) [1],
which are then used to suggest the better product on e-commerce websites. The user-selected products will
receive recommendations from this system. The user can prefer the same earlier based on ideas and
preferences [2]. Due to the effects of widespread web-based social networking use, this new area of research
is becoming more and more important. The majority of the existing structures and resources are made
specifically for English or other European languages. There are not many resources available for Arabic,
despite the fact that it is currently one of the top 10 most used languages online.

This activity primarily focuses on the task of recommending various things to Algerian clients based on the
evaluations of other clients. Our focus is on the top four languages spoken in Algeria: Arabic, Algerian slang,
French, and English. Examining the conclusions is concerned with the emotions and attitudes expressed in a
text. Due to the widespread use of the internet and online social media to convey an extremely large number
of theories, it is expanding swiftly right now. In keeping with this, we created a paradigm for extremities
discovery that has proven productive in prior trial studies [1-4]. This framework changes audits messages into
a numerical assessment and feeds them into a suggestion framework to actualize a community.

Separating. Clients' feelings are expressed unequivocally with a vote or certainly with remarks. Those remarks
are written in normal language with explicit jargon. They have an extremity score to anticipate the vote related
with this remark and make suggestions about things that may intrigue different clients.

By combining the domains of assumption inquiry and proposal with community-oriented sifting to produce a
unique and functional recommender system, this study is peculiar. We suggest integrating a multilingual
suggestion framework with a semi-administered grouping-based assessments inquiry methodology. The
process of suppositions arrangement immediately distinguishes the collection of factual highlights, such as
word count, emotionalism, tendency to, reflexivity, etc. Additionally, the next highlights vector will be a
numerical representation of the survey's content created by the semi-managed SVM during the

546
Special Issue on Innovative Research in Management, Applied Science and Engineering
Stochastic Modeling & Applications ISSN: 0972-3641
Vol. 26 No. 3 (January – June, Special Issue 2022 Part - 8) UGC CARE APPROVED JOURNAL

characterization stage. The choice in favour of the community separating from the recommendation stage is
finally calculated using an extreme score.

Sentiment Analysis (SA) is mostly used now a days in many applications to overcome the various issues in
recommended systems. SA is done on different types of movies, products and various social marketing
techniques [3]. SA is most complicated and complex task. This will build the lexicon of words with positive
and negative variations, and finds the different types of users views by analyzing the words in the text with the
lexicon [4]. At times this feeling is delegated positive negative and impartial. Notion investigation gives the
conclusions about the data so that to settle on choices in different areas. Conclusion investigation is one of the
well known procedures in regular language preparing, which extricate abstract data from the offered
information and group thoughts. The way toward grouping estimation extremity will exist in three degrees of
examination dependent on the given information.

Report level: This degree of feeling grouping thinks about the whole archive assessment and predicts the
assessment of the record as either positive or negative.

Sentence level: This degree of slant grouping thinks about the sentence assessment and predicts the
assessment of the sentence as positive or negative or unbiased. On the off chance that the given sentence is
emotional, this degree of order predicts the assessment of the sentence as positive or negative.

Angle level: This level of conclusion order will describe the hypothesis regarding the elements' explicit
sections. The findings or feelings at each level will be transmitted through this degree of feeling analysis.
Slant examination assignments include a variety of techniques that are arranged mostly into two types of
approaches, like as

AI approach: The AI approach was arranged into two procedures, one is directed learning and other is solo
learning. Regulated learning strategy will foresee the extremity of the objective information or test
information dependent on the preparation dataset with a limited arrangement of classes, for example, positive
and negative. While unaided learning procedure is proposed when there doesn't exist a likelihood to give
earlier preparing dataset to mine information.

2. LITERATURE REVIEW
2.1 Sentiment Analysis
SA is a tactic that falls within the category of everyday language processing. Estimation investigation, often
known as assessment mining, is the process of identifying human emotions and beliefs. According to the
degree of arrangement on a particular record or sentence, it also ranks whether the given content is positive,
negative, or occasionally impartial. There are various strategies for completing the slant examination task. By
identifying the idea or evaluation of the abstract element within a content, this assignment is completed. The
methods used to categorise a piece of writing depend on the emotions it conveys, such as whether they are
neutral, neutral, or positive.

547
Special Issue on Innovative Research in Management, Applied Science and Engineering
Stochastic Modeling & Applications ISSN: 0972-3641
Vol. 26 No. 3 (January – June, Special Issue 2022 Part - 8) UGC CARE APPROVED JOURNAL

Figure 2.1: Sentiment Classification Techniques.

The assessment examination is cultivated by ordering the strategies into AI and dictionary based
methodology. Again the AI approach is characterized into administered and solo AI methods. Under
administered realizing there exists chiefly Support Vector Machine (SVM), Neural Networks (NN), Naive
Bayes (NB), Maximum Entropy (ME) approaches.

Support Vector Machines (SVM): It is a direct classifier for example a non-probabilistic classifier. It is an
administered AI procedure so it requires a lot of preparing information. It arranges the preparation information
by a hyper-plane, which goes about as a limit to isolate the preparation information dependent on
classifications. In view of limits the first items or test information is planned to the grouped preparing
information, which is called change. After the change, the planned articles are directly distinct.

Neural Networks (NN): It additionally goes under direct classifier, which predominantly goes under a
regulated AI method. The direct classifier under this strategy is calculated relapse, which can be seen as a
solitary layer neural system, used to prepare the prescient model.

Navie Bayes (NB): This classifier is a well known probabilistic classifier. It utilizes restrictive likelihood to
characterize words into their separate classifications. The benefit of Naive Bayes classifier is, it needs little
dataset for preparing. The contingent likelihood of given test information point from n expressions of the
preparation set is given by:

Maximum Entropy (ME): It is additionally a probabilistic classifier which is likewise called exponential
classifier or log-straight classifier. This classifier extricates some arrangement of highlights from the
information, consolidating them directly and utilizing this entirety as a type. This area restricted to the writing
review for the most part on AI draws near and various levels and sorts of notion examination procedures. The
proposed marvel was to perform assessment investigation of the gathered tweets, which go about as a corpus.
By utilizing this wonder can have the option to decide positive, negative and unbiased assessments for a
report? The principle commitments to propose this method was to perform measurable etymological

548
Special Issue on Innovative Research in Management, Applied Science and Engineering
Stochastic Modeling & Applications ISSN: 0972-3641
Vol. 26 No. 3 (January – June, Special Issue 2022 Part - 8) UGC CARE APPROVED JOURNAL

examination of the gathered corpus with positive and negative sentences and manufacture a supposition
grouping by utilizing this gathered corpus. The test assessment was led on constant small scale blog entries
utilizing the assumption grouping framework. Every one of these procedures will be directed consequently
without human exertion. This proposed technique for the most part beats on unigrams contrasted with bigram
by utilizing Naive Bayes approach. This methodology orders the 3-class grouping, for example it decides the
estimation extremity of the given objective information as positive, negative or impartial. This proposed
strategy additionally applied to other assessment classifiers, for example, SVM and CRF. For include
extraction, the creator utilizes AI procedure, for example, Bag of Words technique, which is a customary
component extraction approach. For leading trial assessment, the creator didn't make reference to the space of
the preparation information and target miniaturized scale blog continuous information. They played out this
regulated AI method on Twitter information. The creator performed k-overlap cross-approval to test the
calculation execution.

3. ENHANCED DEEP LEARNING FOR RECOMMENDER SYSTEM

Deep learning has gotten all the more impressive due the measure of accessible preparing information has
enormous development. The pattern of the utilization of profound learning innovation was additionally
impacted by the expansion in the capacity of equipment and framework that certified to perform processing.
Late 5 years, profound learning has capacity to increment critical execution in a few viewpoints like picture
preparing, text mining, regular language handling, voice perceiving. Numerous specialists going to lead
research in recommender framework field include profound learning innovation. The world exploration
association for instance ACM as a significant examination in recommender framework field welcomed for
individuals in around the globe who intrigued by this exploration territory to included.

In DL, various traditional and dynamic algorithms are present such as DBN, Convolutional Neural Network
(CNN) [7], Recurrent Neural Network (RNN) [8], Deep Autoencoder [9]. DBN, is one of the deep generative
models, this consists of more number of Restricted Boltzmann Machines (RBMs) [10] which is random neural
network that consists of visual units for single layer and hidden units for one layer.

Oh et al. [13] explained a different technique, which is based on DBM, this will analyze the user preference
for the recommender system for the personal news. The author Zahalka et al. [12] explained about the
interactive and multi-modal content-based venue explorer, which is integrated with CNN to create the optical
aspects. Lin et al. [13] created the hierarchical deep CNN system for garments result in recommender system.
Cui et al. [14] introduced the recommender system with video that merges the DBN with CF. The author Lei
et al. [15] explained about the image recommendations based on the two way deep network models, that
contains two CNNs and a total-connection neural network. The author Chiliguano et al. [16] implemented the
recommender system with CNN which is based on music. Wu et al. [17] designed the recommender system
for E-Commerce sites with a deep RNN. The author Zuo et al. [18] introduced the system with a tag-aware
that combined Sparse Auto-encoder with user-based CF. The author Unger et al. [19] explained a
recommender system which is context-aware by using Auto-encoder to analyze possible context information.
The author Deng et al. [20] designed recommendation system which is trust-based by using Deep Auto-
encoder in social networks. Sivaparvathi V. et al [21] introduces the deep learning sentiment analysis based on
Intelligent Product Recommendation System which is used improve the classification accuracy to produce
correct recommendations on various products.

3.1. Single and Hybridization Approach in Deep Learning

In Enhanced DL, the utilization of DL has a few progressively development so as to accomplish the best
execution, there are a few scientists who have an objective to upgrade with single profound learning and
hybridization profound getting the hang of, agreeing analyst to utilize the term composite DL.

549
Special Issue on Innovative Research in Management, Applied Science and Engineering
Stochastic Modeling & Applications ISSN: 0972-3641
Vol. 26 No. 3 (January – June, Special Issue 2022 Part - 8) UGC CARE APPROVED JOURNAL

Utilizing single DL strategy and deep composite model (recommender framework which includes at least two
DL strategies).

The MLP can easily simulate the asymmetrical interactions between clients and objects; CNN is capable of
extracting local and global representations from diverse information sources, such as printed and visual data.
RNN enables the recommender framework to show the ephemeral parts of rating information and continuous
effects of substance data.
• Deep Composite Model. Some profound learning based proposal models use more than one profound
learning method. The inspiration is that diverse profound learning strategies can supplement each other and
empower an all the more impressive cross breed model. There are numerous potential mixes of these eight
profound learning procedures however not all have been misused.

4. DEEP LEARNING MODELS FOR SENTIMENT ANALYSIS WITH RECOMMENDER SYSTEM

4.1 Recommender System Framework
Recommender frameworks are utilized to assemble suggestions by preparing data from effectively
accumulated fluctuated sort of information; recommender frameworks bolster the clients to settle on choices
with respect to their inclinations about the item or thing. Notion investigation assumes a significant job in
building a recommender framework. Utilization of supposition examination strategy on the cloud stage filled
in as a choice emotionally supportive network for the recommender framework on the cloud.

The crucial thought behind this system is to gather data from hubs. The cloud stage and interpersonal
organization that is utilized to gather data from hubs can be twitter, Facebook and so on. The gathered data
from hubs incorporates scope and longitude geographic area. Utilization of opinion investigation strategy on
the gathered data utilizing distinctive profound learning models, for example, multilayer perceptron model,
convolutional neural system model and recursive neural system model. The structure that is utilized for
recommender framework dependent on profound learning models on feeling examination was appeared in
figure 4.1 and the neural system models that are considered as profound learning models were appeared in
figure 4.2 separately.

Neural systems are little registering units with interconnected processors. The neural system models are
utilized to tackle characteristic language preparing issues, for example, learning word installing and
supposition arrangement. The assumption grouping was performed on the content just as on picture. The
profound learning models for supposition grouping were ordered in to take care of forward neural system
models and recursive neural system model. The feed-forward neural system models were again ordered into
two models, for example, multilayer perceptron model and convolutional neural system models. The entirety
of the three models were applied on text for conclusion arrangement, where as a convolutional neural system
can be utilized to group pictures. So this neural system model can be applied on the content just as on picture
too.

550
Special Issue on Innovative Research in Management, Applied Science and Engineering
Stochastic Modeling & Applications ISSN: 0972-3641
Vol. 26 No. 3 (January – June, Special Issue 2022 Part - 8) UGC CARE APPROVED JOURNAL

Figure 4.1: Recommender System Framework.

Figure 4.2: Deep learning Models.

4.2 Recursive Neural Network (RNN)

RNN is a different type of neural network which can be seen as a generalization of the recurrent neural
network. The recursive neural network is used to learn a tree structure from data. Recursive neural networks
have effectively applied to process information structures as contributions to neural nets in characteristic
language handling. Given the auxiliary portrayal of a sentence, recursive neural systems can recursively create
parent portrayals in a base up style, by joining tokens to deliver portrayals for phrases, inevitably the entire
sentence.

In recursive neural systems, the highlights of a sentence are spoken to in a high-dimensional vector space after
being eliminated, and then the highlights are presented to the neural system. Recursive neural networks are
capable of efficiently comprehending sentence structure. This neural system model is therefore practical and
plausible for examining opinions. Consider the phrase or archive Di is "doing amazing work" for the purposes
of demarcation. The provided report is a jumble of words. Highlights are taken out of the text and spoken to in
a high-dimensional vector space. The final order for a particular information succession might then be created
using the sentence level depiction of the given archive, as shown in figure 4.3

551
Special Issue on Innovative Research in Management, Applied Science and Engineering
Stochastic Modeling & Applications ISSN: 0972-3641
Vol. 26 No. 3 (January – June, Special Issue 2022 Part - 8) UGC CARE APPROVED JOURNAL

Figure 4.3: Sentiment Classification Using Recursive Neural Network

5. SYSTEM MODEL
Recommender System is a method, which gives proposals to things to be useful to a client. The methods that
include in recommender framework incorporate data recovery which is an example of an information mining
procedure. This procedure additionally includes different fields, for example, Artificial Intelligence (AI),
Information Technology (IT), Statistics, Decision Support System (DSS) and so forth. In this part, the
information mining procedure incorporates text mining dependent on the question of the client that removes
data from the cloud stage, for example, Twitter API. These removed tweets dependent on geo-location. The
tweets that are separated will be pre-processed and perform slant investigation. The feeling examination
strategy gives slant and assessments in regards to the responses to the questions gave by the client on the
cloud. These assessments and conclusions will go about as suggestions to the client. In light of the gave
assessments and sentiments the client can choose whether the individual needed to pick the item or not. The
principle key pretended strategy is notion investigation, which goes about as the Decision Support System. In
light of the supposition investigation procedure we can outwardly characterize the conclusions and notions.
The model was obviously spoken to in figure5.1 shows the connection between the cloud and the learning
framework.

Figure 5.1: Model that shows the Interaction between the Cloud and the Learning System.

DLS Framework:
Information gathered from nodes is what helps this system the most. Nodes provide knowledge that includes
geographic information such as latitude and longitude. Apply thorough emotional research after considering
the information acquired. We react to the feedback either positively or negatively. The solution is predicated

552
Special Issue on Innovative Research in Management, Applied Science and Engineering
Stochastic Modeling & Applications ISSN: 0972-3641
Vol. 26 No. 3 (January – June, Special Issue 2022 Part - 8) UGC CARE APPROVED JOURNAL

on the previous history of the user. Based on prior experiences, the consumer's thoughts can be predicted. We
determine if a positive or neutral analysis of the product is favourable based on our feelings. Twitter,
Facebook, and other social networks can be utilised as a cloud platform and social network to gather data from
nodes.

Figure 5.2: DLS Framework.

Deep Learning Model:

This area proposes diverse profound learning models for include extraction and characterization of learning
information and furthermore to utilize profound learning model to anticipate the item reaction either positive
or negative or nonpartisan. This section talks about various models for include age and notion
characterization. Utilization of profound learning models to the assumption investigation procedure will
upgrade the presentation. Profound learning models incorporate direct and non-straight classifiers with high
dimensional scanty vector or low dimensional thick vector.

def sentimentanalysis(arr, n):

poswords p[ ]:
negwords n[ ]:
neuwords neu[ ]:
v=[]
# For each element of array.
for i in range(n):
# Try to find the negative analysis
# of arr[i] from i + 1 to n
for j in range( i + 1,n) :
# If absolute values are
# equal print pair.

553
Special Issue on Innovative Research in Management, Applied Science and Engineering
Stochastic Modeling & Applications ISSN: 0972-3641
Vol. 26 No. 3 (January – June, Special Issue 2022 Part - 8) UGC CARE APPROVED JOURNAL

if (abs(arr[i]) == abs(arr[j])) :
v.append(abs(arr[i]))
# If size of vector is 0, therefore
# there is no element with positive analysis
# negative analysis, print "0"
if (len(v) == 0):
return;
# Sort the vector
v.sort()
for i in range(len( v)):
print(-v[i], "" , v[i], end = " ")
if __name__ == "__main__":
arr = { poswords p[]:, negwords n[]:, neuwords neu[]: }
n = len(arr)
printPairs(arr, n)

Performance Evolution
By utilizing the performance measures namely False Positive Rate, False Negative Rate, Sensitivity,
Specificity and Accuracy, the performance of the system is estimated. The basic count values such as True
Positive (TP), True Negative (TN), False Positive (FP) and False Negative (FN) are used by these measures.
False Positive Rate (FPR)
The percentage of cases where a data was classified to normal text, but in fact it did not.
𝐹𝑃
𝐹𝑃𝑅 =
𝐹𝑃 + 𝑇𝑁
False Negative Rate (FNR)
The percentage of cases where a data was classified to abnormal text, but in fact it did.
𝐹𝑁
𝐹𝑁𝑅 =
𝐹𝑁 + 𝑇𝑁
Sensitivity
The proportion of actual positives which are correctly identified is the measure of the sensitivity. It relates to
the ability of the test to identify positive results.
𝑁𝑜. 𝑜𝑓 𝑇𝑃
𝑆𝑒𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 =
𝑁𝑜. 𝑜𝑓 𝑇𝑃 + 𝑁𝑜. 𝑜𝑓 𝑇𝑁
Specificity
The proportion of negatives which are correctly identified is the measure of the specificity. It relates to the
ability of the test to identify negative results.
𝑁𝑜. 𝑜𝑓 𝑇𝑁
𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 =
𝑁𝑜. 𝑜𝑓 𝑇𝑁 + 𝑁𝑜. 𝑜𝑓 𝐹𝑃
Accuracy: This will calculate the overall accuracy of the clusters.
𝑇𝑃 + 𝑇𝑁
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁

554
Special Issue on Innovative Research in Management, Applied Science and Engineering
Stochastic Modeling & Applications ISSN: 0972-3641
Vol. 26 No. 3 (January – June, Special Issue 2022 Part - 8) UGC CARE APPROVED JOURNAL

Experimental Results
In this paper, the synthetic E-commerce Application with product review dataset is utilized for sentiment
analysis. Python programming is used to process the dataset. The system configuration is higher with 8 GB
Ram and 1 TB hard disk with I7 processor for fast response. For review analysis, 10 products that are mostly
popular products belongs to electronics such as mobiles, televisions, laptops and other house hold products
are selected and every product has more than 50 reviews and 8 attributes for every product. The sentiment
analysis is classified into three categories such as positive, negative and neutral. The proposed system DLS
applied on this dataset.
Table 1 shows the 10 most products with review analysis.

Product Name Total Positive Negative Neutral

Number of reviews reviews reviews
Reviews
Samsung Galaxy M11 154 99 15 40

Apple iPhone 8 Plus 131 111 10 10

OPPO A9 2020 122 88 13 21

Mi 4A PRO 76 46 20 10
Nokia 139cm (55 inch) Ultra HD (4K) 67 47 8 12

Voltas 1.5 Ton 3 Star Split Inverter AC 181 121 41 17

Samsung 253 L Frost Free Double Door 3 199 136 38 23

Star
OPPO F9 Pro 134 88 20 24

Motorola G8 Power Lite 156 111 26 17

Vivo U10 171 141 23 6

Table 1 Results Analysis

Algorithms Sensitivity Specificity Accuracy

SVM 89.7 88.1 91.2
DLS 97.7 98.2 99.6

Table 2 Performance of SVM and DLS with the selected E-commerce sentiment dataset.

555
Special Issue on Innovative Research in Management, Applied Science and Engineering
Stochastic Modeling & Applications ISSN: 0972-3641
Vol. 26 No. 3 (January – June, Special Issue 2022 Part - 8) UGC CARE APPROVED JOURNAL

Figure 1: performance metrics

CONCLUSION
Although deep learning machine methods excel in some areas, such as photo processing, common language
preparation, and discourse perception, deep learning machine methods used in recommender systems do not
always perform the most difficult tasks. The typical recommender framework tends to recommend items
based on ratings. In this way, the intensity of the available ratings are insufficient to assess the idle factor
between customer and item due to virus start and information issue. This problem needs to be resolved, and
one arrangement that might go along with collective splitting is content-based. Deep learning plays a large
role in this situation to remove assistant data assets for potential attempts to add data, such as audit messages,
photographs, client profiles, sound, and so on. can then be put to use. This analysis suggests a structure for
opinion investigation with a recommender framework that is based on multiple profound learning models. The
controlled and solo learning methods are the foundation of the proposed profound learning models. For
assessment, a DLS system that applies hypothesis analysis to more limited blog literary data using sham
variables, a multilayer perceptron model that makes use of extensive learning, and a summed up straight
model known as computed relapse was developed.

References
[1] Ricci, Francesco, Lior Rokach, and Bracha Shapira. Introduction to recommender systems
handbook.springer US, 2011.
[2] Cantador, Ivan, Alejandro Bellogín, and Pablo Castells. "A multilayer ontology-based hybrid
recommendation model." AI Communications 21.2-3 (2008): 203-210.
[3] Pang, Bo, Lillian Lee, and Shivakumar Vaithyanathan. "Thumbs up?: sentiment classification using
machine learning techniques." Proceedings of the ACL-02 conference on Empirical methods in natural
language processing-Volume 10. Association for Computational Linguistics, 2002.
[4] Ye, Qiang, Ziqiong Zhang, and Rob Law. "Sentiment classification of online reviews to travel destinations
by supervised machine learning approaches." ESA 36.3 (2009): 6527-6535.
[5] Pang, Bo, and Lillian Lee. "Opinion mining and sentiment analysis." Foundations and Trends in
Information Retrieval 2.1–2 (2008): 1-135.
[6] Krishna, P. V., Misra, S., Joshi, D., & Obaidat, M. S. (2013, May). "Learning automata based sentiment
analysis for recommender system on cloud." CITS, 2013 International Conference on (pp. 1-5). IEEE.
[7] Lawrence S, Giles C L, Tsoi A C, et al. Face recognition: a convolutional neural-network approach.[J].
IEEE Transactions on Neural Networks, 1997, 8(1):98-113.
[8] Martens J, Sutskever I. Learning recurrent neural networks with hessian-free optimization[C],
International Conference on Machine Learning, ICML 2011, Bellevue, Washington, Usa, June 28 - July.
DBLP, 2012:1033-1040.
[9] Hinton G E, Salakhutdinov R R. Reducing the dimensionality of data with neural networks [J]. Science,
2006, 313(5786):504.
[10] Fischer A, Igel C. An Introduction to restricted boltzmann machines[C], Iberoamerican Congress on
Pattern Recognition. Springer, Berlin, Heidelberg, 2012:14-36.

556
Special Issue on Innovative Research in Management, Applied Science and Engineering
Stochastic Modeling & Applications ISSN: 0972-3641
Vol. 26 No. 3 (January – June, Special Issue 2022 Part - 8) UGC CARE APPROVED JOURNAL

[11] Oh K J, Lee W J, Lim C G, et al. Personalized news recommendation using classified keywords to
capture user preference[C]// International Conference on Advanced Communication Technology. IEEE,
2014:1283-1287.
[12] Zahálka J, Rudinac S, Worring M. Interactive multimodal learning for venue recommendation [J]. IEEE
Transactions on Multimedia, 2015, 17(13):2235-2244.
[13] Lin K, Yang H F, Liu K H, et al. Rapid clothing retrieval via deep learning of binary codes and
hierarchical search[J]. 2015:499-502.
[14] Cui H, Qin X. The Video Recommendation system based on DBN[C], IEEE International Conference on
Computer and Information Technology; Ubiquitous Computing and Communications; Dependable,
Autonomic and Secure Computing; Pervasive Intelligence and Computing. IEEE, 2015:1016-1021.
[15] Lei C, Liu D, Li W, et al. Comparative deep learning of hybrid representations for image
recommendations[C], IEEE Conference on Computer Vision and Pattern Recognition. IEEE, 2016:2545-
2553.
[16] Chiliguano P, Fazekas G. Hybrid music recommender using content-based and social information[C]//
IEEE International Conference on Acoustics, Speech and Signal Processing. IEEE, 2016:2618-2622.
[17] Wu S, Ren W, Yu C, et al. Personal recommendation using deep recurrent neural networks in
NetEase[C]// IEEE, International Conference on Data Engineering. IEEE, 2016:1218-1229.
[18] Zuo Y, Zeng J, Gong M, et al. Tag-aware recommender systems based on deep neural networks [J].
Neurocomputing, 2016, 204:51-60.
[19] Unger M, Bar A, Shapira B, et al. Towards latent context-aware recommendation systems [J].
Knowledge-Based Systems, 2016, 104:165-178.
[20] Deng S, Huang L, Xu G, et al. On deep learning for trust-aware recommendations in social networks [J].
IEEE Transactions on Neural Networks & Learning Systems, 2016, 28(5):1164-1177.
[21] Sivaparvathi V., Lavanya Devi G., Rao K.S. (2020) A Deep Learning Sentiment Primarily Based
Intelligent Product Recommendation System. In: Kumar A., Paprzycki M., Gunjan V. (eds)
ICDSMLA 2019. Lecture Notes in Electrical Engineering, vol 601. Springer, Singapore.[22]

557
Special Issue on Innovative Research in Management, Applied Science and Engineering