A Local Differential Privacy based Hybrid Recommendation Model with

BERT and Matrix Factorization

Jeyamohan Neera

, Xiaomin Chen

, Nauman Aslam, Biju Issac and Eve O’Brien

Department of Computer and Information Sciences, Northumbria University, U.K.

Keywords:

Local Differential Privacy, Sentiment Analysis, Collaborative Filtering, Privacy.

Abstract:

Many works have proposed integrating sentiment analysis with collaborative ﬁltering algorithms to improve

the accuracy of recommendation systems. As a result, service providers collect both reviews and ratings,

which is increasingly causing privacy concerns among users. Several works have used the Local Differential

Privacy (LDP) based input perturbation mechanism to address privacy concerns related to the aggregation of

ratings. However, researchers have failed to address whether perturbing just ratings can protect the privacy

of users when both reviews and ratings are collected. We answer this question in this paper by applying an

LDP based perturbation mechanism in a recommendation system that integrates collaborative ﬁltering with a

sentiment analysis model. On the user-side, we use the Bounded Laplace mechanism (BLP) as the input rating

perturbation method and Bidirectional Encoder Representations from Transformers (BERT) to tokenize the

reviews. At the service provider’s side, we use Matrix Factorization (MF) with Mixture of Gaussian (MoG) as

our collaborative ﬁltering algorithm and Convolutional Neural Network (CNN) as the sentiment classiﬁcation

model. We demonstrate that our proposed recommendation system model produces adequate recommendation

accuracy under strong privacy protection using Amazon’s review and rating datasets.

1 INTRODUCTION

Collaborative Filtering (CF) based recommendation

systems predict a user’s preference using similar users

or relevant items. Even though CF algorithms pro-

duce satisfactory recommendation accuracy to an ex-

tent, they build upon the presumption that ratings re-

ﬂect a user’s true preferences and the actual quality

of an item. This presumption does not always cor-

respond with real scenarios, as impugn users might

give low ratings and tolerant users tend to give high

ratings. Such hypotheses play a huge role when a

customer decides whether an item is suitable or un-

suitable to fulﬁl their needs (Raghavan et al., 2012).

Besides, users who give similar ratings to an item

may have experienced distinct degrees of satisfaction

(Cheng et al., 2018). Finally, CF algorithms often suf-

fer from data sparseness problems due to a lack of rat-

ings which results in low recommendation accuracy

(Mobasher et al., 2007). Therefore, CF algorithms

have started incorporating sentiment analysis to ad-

dress these issues. Sentiment analysis is a technique

https://orcid.org/0000-0001-8771-4193

https://orcid.org/0000-0001-9267-355X

used to categorize text-based data to better understand

users’ attitudes and opinions in several domains. By

combining sentiment analysis with a CF algorithm,

the service provider can use ratings and reviews to

further improve the recommendation accuracy. How-

ever, collecting both ratings and reviews imposes a

higher risk of causing a violation of user privacy.

This paper proposes applying a Local Differen-

tial Privacy (LDP) based perturbation mechanism to

a recommendation system that combines sentiment

analysis with a CF algorithm to predict a user’s pref-

erences. We perturb the user’s original ratings lo-

cally using a Bounded Laplace input perturbation

mechanism (BLP) before sending them to the service

provider. A deep learning-based sentiment analysis

model is used to analyze user reviews. However, we

propose that user reviews are tokenized locally and

then sent to the service provider for aggregation and

classiﬁcation purposes. Since the service providers

only aggregate the tokenized reviews, they cannot in-

fer sensitive information about users without access to

the original review data. Additionally, the perturbed

ratings are used as the input sentiment labels, prevent-

ing the service provider from learning a user’s actual

sentiment using their tokenized review data. We use

Neera, J., Chen, X., Aslam, N., Issac, B. and O’Brien, E.

A Local Differential Privacy based Hybrid Recommendation Model with BERT and Matrix Factorization.

DOI: 10.5220/0011266800003283

In Proceedings of the 19th International Conference on Security and Cryptography (SECRYPT 2022), pages 325-332

ISBN: 978-989-758-590-6; ISSN: 2184-7711

325

Matrix Factorization (MF) with Mixture of Gaussian

(MoG) as our CF algorithm. The MF-MoG model

that runs at the service provider’s end estimates the

noise added to the aggregated perturbed ratings and

predicts missing ratings simultaneously. The results

of the empirical study show that our proposed recom-

mendation model signiﬁcantly improves the recom-

mendation accuracy under a strong privacy guarantee.

2 RELATED WORK

Many works (Paterek, 2007; Pal et al., 2017) have

investigated different approaches to improve the pre-

dictive performance of CF algorithms. Wang et al.

(Wang et al., 2018) incorporated a CF recommenda-

tion system with a sentiment analysis model to obtain

an optimised preliminary recommendation list ﬁrst

and then used it to produce a ﬁnal recommendation

list at the end. In another work, Osman et al. (Os-

man et al., 2019) proposed a recommendation system

where the context of user’s comments was taken into

consideration and used to produce recommendations.

Such approaches are more suitable when ratings are

sparse and aid immensely in increasing recommenda-

tion accuracy.

Even though these proposed solutions improve

the recommendation accuracy, they also cause se-

rious privacy concerns in recommendation systems,

and several works have proposed new ways to tackle

these privacy concerns. Li and Sarkar (Li and Sarkar,

2011) presented an encryption-based solution to ad-

dress the privacy concerns in a user-based CF recom-

mendation system. McSherry and Mironov (McSh-

erry and Mironov, 2009) proposed a differential pri-

vacy based solution in which noise is added to the

item-to-item co-variance matrix. They proved that

using differential privacy in recommendation systems

offered strong privacy protection to users through ex-

periments. Wang and Duan (Wang et al., 2016) pro-

posed a privacy quantiﬁcation model which synthe-

sised an individual’s inﬂuence and system privacy fac-

tors based on a user’s perception. Similarly, Sutanto

et al. (Sutanto et al., 2013) designed a personalised,

privacy-safe application that enabled users to control

their privacy. However, this method failed to produce

accurate recommendations.

The aforementioned works consider the service

providers to be trustworthy. Considering the existence

of untrustworthy service providers, many works have

begun to investigate the application of LDP in recom-

mendation systems. LDP perturbs users’ data on the

user-side rather than on the service provider side. Liu

et al. (Liu et al., 2015) proposed a recommendation

system where noise is added to users’ ratings locally

on the user-side through a randomised perturbation

method. Meng et al. (Meng et al., 2018) in their work

proposed using LDP based input perturbation mech-

anism only on ratings that were considered sensitive

before sending them to the service provider. Shen and

Jin (Shen and Jin, 2014) aimed to hide users’ prefer-

ence toward an item using an instance-based admis-

sible mechanism on all the ratings. Shin et al. (Shin

et al., 2018) proposed an LDP-based recommendation

model which used a randomised response mechanism

to add noise to users’ ratings.

Most solutions that have been proposed in the lit-

erature to address privacy concerns in recommenda-

tion systems concentrate only on ratings based rec-

ommendation systems and have not yet addressed pri-

vacy concerns when user reviews are taken into con-

sideration. Therefore, we propose using an LDP per-

turbation mechanism that perturbs users’ ratings and

tokenize reviews on the user-side to protect user’s pri-

vacy from the service provider. We also propose esti-

mating the noise added to the ratings at the server-side

to improve the recommendation accuracy.

3 PRELIMINARIES

3.1 Local Differential Privacy

Each user perturbs their data locally before sending it

to the service provider in the LDP setting. So the ser-

vice provider collects only the perturbed data instead

of the original data. This approach can signiﬁcantly

protect users’ privacy from an untrustworthy service

provider. Intuitively in LDP based settings, the data

aggregator cannot infer whether a user’s input x or x

produces the output y. Therefore, LDP offers plausi-

ble deniability to users.

Deﬁnition 1. A randomised mechanism M satisﬁes

ε-LDP if for all possible pairs of user input x, x

and

any subset y of all possible outcomes, we have the

following inequality:

Pr[M(x) ∈ y] ≤ e

× Pr[M(x

) ∈ y].

The privacy budget ε acts as a metric of privacy

loss at a perturbed data.

3.2 Bounded Laplace Mechanism

BLP perturbs data by ignoring any output values that

do not fall inside a predeﬁned domain and re-samples

the noise from a Laplace distribution until the output

value lies within the given bound. Unlike the Laplace

SECRYPT 2022 - 19th International Conference on Security and Cryptography

326

mechanism, BLP sanitises the perturbed output with

bounding constraints (Holohan et al., 2018).

Deﬁnition 2. (Bounded Laplace Mechanism) Given

a scale parameter b and a domain rating interval of

(l, u), the Bounded Laplace mechanism M

BLP

: R →

∗

is given by a conditional probability density func-

tion as follows:

∗

) =

(

(b)

−

∗

−r|

, if r

∗

∈ [l,u],

0, if r

∗

/∈ [l,u],

where C

(b) is a normalisation constant, r is the input

to the BLP and r

∗

is the perturbed output.

The normalisation constant C

(b) is given as:

(b) = 1 −



exp(−

r − l

) + exp(−

u − r

)



3.3 Bidirectional Encoder

Representations from Transformers

The objective of sentiment analysis is to determine

whether a user’s review communicates their posi-

tive or negative opinion. Deep learning-based tech-

niques are proven to be highly effective in identifying

user sentiments in several applications (Goularas and

Kamis, 2019; Zarzour et al., 2021). Hence, we use

BERT (Bidirectional Encoder Representations from

Transformers) to create word embedding in our senti-

ment analysis model. BERT is a natural language pro-

cessing model which was introduced by the Google

AI team in 2018 (Devlin et al., 2018). BERT provides

a contextualised representation, unlike other word-

embedding models such as Word2Vec and GloVe, that

generates a single representation for each word in a

given input text. BERT uses transformer encoder lay-

ers to learn these contextual relations between words

in a given text. The training of the BERT model takes

place in two stages: pre-training and ﬁne-tuning.

The BERT model is trained on an unlabelled cor-

pus that contains text from English Wikipedia and

Book Corpus dataset in the pre-training stage. These

large collections of words allow BERT to capture

extensive language knowledge. The resulting pre-

trained model can then be ﬁne-tuned for speciﬁc

NLP tasks such as sentiment analysis. The ﬁne-

tuning stage is an essential step in training the BERT

model. Even though pre-training produces a bidirec-

tional unsupervised language representation of texts,

ﬁne-tuning allows this representation to be used in

any NLP related tasks. In the ﬁne-tuning stage, the

BERT model is initialised with the same parameters

as the pre-trained model, and the parameters are then

adjusted according to the labelled data. BERT model

has several deployment types based on their model

conﬁgurations, such as RoBERTa (Liu et al., 2019)

which proposed a method to improve the training pro-

cess of BERT and ALBERT (Lan et al., 2019) which

reduces the model size through parameter sharing and

factorising techniques.

4 SYSTEM DESIGN

In this section, we describe our proposed LDP based

recommendation system that combines a CF algo-

rithm with a sentiment analysis model and uses BLP

as the input perturbation mechanism. The aim is to

improve the recommendation accuracy while provid-

ing privacy protection to the users. Fig.1 illustrates

the architecture of the proposed recommendation sys-

tem. We assume that users can report their actual

ratings and reviews anonymously so that the service

provider can display these anonymous reviews and

ratings on their platform. We do not discuss the ar-

chitecture required for anonymous reporting as it is

beyond the scope of this paper. We use the MF-MoG

model as the CF algorithm that uses perturbed ratings

as the input. The predicted rating combination mod-

ule uses the outputs of the CNN classiﬁcation model

and the MF-MoG model to produce the ﬁnal predicted

rating.

Algorithm 1: BLP Mechanism for Noise Sampling.

1: Input to the Mechanism: Original Rating

2: Output of the Mechanism: Perturbed Rating

3: Generate noise from the Laplace distribution with

mean 0 and variance of b

4: Perturbed rating = Original rating + noise

5: If Perturbed rating is in given domain interval:

6: Perturbed rating is set

7: else

8: repeat Step 3

9: Return Perturbed rating to SP

4.1 Input Rating Perturbation at

User-side

First, users use BLP as an LDP perturbation mecha-

nism to perturb their original ratings. The perturbed

ratings are then sent to the service provider for aggre-

gation. Algorithm 1 describes how a perturbed rating

is generated using the BLP mechanism. It has been

proven (Neera et al., 2021) that a sufﬁcient condition

for BLP to satisfy ε-local differential privacy in rec-

ommendation systems is when the local sensitivity is

∆ f = l − u where l is minimum and u is the maxi-

mum rating in a given rating scale. The BLP mech-

A Local Differential Privacy based Hybrid Recommendation Model with BERT and Matrix Factorization

327

MF WITH

MoG

PREDICTED

RATING

COMBINATION

BLP

(Privacy

Budget ε)

BERT TOKENIZER

PRE-PROCESSING

DATA

AGGREGATION

𝑃

𝑀𝐹

𝑃

𝑆𝑒𝑛𝑡

𝑃

𝐹𝑖𝑛𝑎𝑙

USER-SIDE

SERVICE PROVIDER SIDE

REVIEWS

RATINGS

PERTURBED RATINGS

DATASET

CNN

CLASSIFICATION

MODEL

REVIEWS AND

PERTURBED RATINGS

DATASET

Figure 1: Proposed LDP-based Recommendation System with Sentiment Analysis and CF Algorithm.

anism ensures that the perturbed output rating is lim-

ited to the rating domain [l,u] and still guarantees that

the adversary cannot infer any information about the

original rating by observing the perturbed rating.

4.2 Review Pre-processing at User-side

Sentiment analysis models require the input review

data to be cleaned and processed before using them in

a classiﬁcation model. The original reviews of users

are pre-processed and converted to a tokenized review

before being sent to the service provider. First, words

that lack relevant information, leading and trailing

spaces, numbers, punctuation and stop words in the

review text are removed. Additionally, the text is

converted to lowercase. Then the cleaned review is

split into individual words and then lemmatized. The

lemmatization process converts the inﬂectional and

derivational forms of a word to its common base form.

For example, the words run, runs, and running are

converted to the base word run. After lemmatization,

we use BERT to compute the sequence embedding.

BERT maps each word into a vector of numerical val-

ues so that words with similar meanings have a sim-

ilar representation. Each user does the review pre-

processing locally and sends only the numerical vec-

tor to the service provider so that the actual review

text is never revealed to the service provider.

4.3 Multi Class Classiﬁcation using

CNN

In our work, we use BERT only as an encoder and

a CNN model as the decoder to conduct sentiment

classiﬁcations. Even though BERT itself can perform

sentiment classiﬁcation, the multi-label classiﬁcation

layer has to be retrained on top of the transformer to

perform sentiment prediction. Fig.2 illustrates the hy-

brid sentiment analysis model used in our recommen-

dation system.

BERT

FEATURE VECTOR

CNN

ReLU

FUNCTION

OUTPUT

Figure 2: Hybrid sentiment analysis.

We use a CNN model for multi-class classiﬁca-

tion at the service provider’s end as it is proven to be

effective in text review classiﬁcation (Salinca, 2017).

Since the CNN model uses perturbed ratings as the

sentiment labels, we hide the true sentiment of the

users from the service provider. This step adds an-

other layer of privacy protection to the sentiment anal-

ysis model and offers plausible deniability to users.

CNN model learns spatial hierarchies of features us-

ing three layers that is convolution, pooling and fully

connected. The ﬁrst two layers do feature extraction,

and the ﬁnal fully connected layer maps the extracted

features into a relevant sentiment. We used a convo-

lution layer with 256 ﬁlters and a linear rectiﬁcation

unit (ReLU) as the activation function. The output of

the CNN model will be the predicted rating P

Sent

4.4 Matrix Factorization with Mixture

of Gaussian Model

We use Matrix factorization with a Mixture of Gaus-

sian model (MF-MoG) (Neera et al., 2021) as the

CF algorithm to make rating predictions on the ser-

SECRYPT 2022 - 19th International Conference on Security and Cryptography

328

vice providers’ side. Since the CF algorithm uses

perturbed ratings as the input, recommendation sys-

tems yield low recommendation accuracy. We use the

Mixture of Gaussian (MoG) on the service provider

side to estimate the noise added to the original rat-

ings to enhance prediction accuracy. Since the post-

processing property of LDP states that any further

processing of a perturbed output of a differentially

private mechanism does not violate the differential

privacy principles (Dwork et al., 2014), the MF-MoG

model still can provide privacy protection to users.

Fig. 3 illustrates the MF-MoG recommendation sys-

tem. The output of the MF-MoG model would be the

predicted rating P

. Algorithm 2 details how the

MF-MoG model estimates noise and predicts missing

ratings.

Algorithm 2: Noise Estimation and Rating Prediction

Model.

1: Input: Perturbed Ratings (R

∗

)

2: Output: User latent factor U and Item latent

factor V

3: Initialisation: Model parameters U,V,Π and Σ

are randomly initialised where Π is the Gaussian

mixture proportion and Σ is the standard devia-

tion.

4: In E-step the posterior responsibility γ

i jk

(x)

is es-

timated

5: For Until convergence

6: (M-Step for updating Σ

(x+1)

and Π

(x+1)

) Model

parameters Σ and Π are computed.

7: (M-Step for estimating V and U) Model param-

eters U and V are updated.

8: (E-step for posterior responsibility γ

i jk

) poste-

rior responsibility γ

i jk

is computed using current

model parameters

9: Return User and Item latent factor matrices U

and V

True

Rating

(r)

Bounded Laplace

Mechanism

User Side

SP Side

Recommendation System

Aggregated

Ratings

MF with

MoG

Top K Recommended Items

Rating Perturbation

Figure 3: LDP based MF recommendation with MoG.

4.5 Predicted Ratings Combination

The predicted ratings combination module combines

outputs from the CNN classiﬁcation model and the

MF-MoG model to produce the ﬁnal predicted rating.

The ﬁnal rating of a user an item is given as:

f inal

= β ∗ P

+ (1 −β) ∗ P

Sent

(1)

where P

f inal

is the ﬁnal predicted rating, P

is the

rating predicted using the MF-MoG model, P

Sent

the rating predicted using sentiment analysis and β is

the parameter that is used to adjust the importance of

each component.

5 EVALUATION

In this section, we present the evaluation results of

our proposed LDP based recommendation model. We

use two datasets, Amazon Toys and Games and Ama-

zon Instant Video, to validate the effectiveness of the

proposed system. We use the Root Mean Squared Er-

ror (RMSE) to evaluate the prediction accuracy and

F-score to evaluate the utility of the recommendation

system.

5.1 Dataset

Table 1 provides a detailed view of the datasets we

used in our evaluation.

Table 1: Datasets.

Dataset Total

Ratings

Total

Reviews

Rating

Scale

Amazon

Toys and

Games

2,252,771 167,597 1 to 5

Amazon

Electronics

583,933 37,126 1 to 5

5.2 Metrics

The privacy budget ε acts as a metric of privacy loss

at perturbed data. The lower the privacy budget ε is

the higher the privacy that is guaranteed. We consider

the value range from 0.1 to 3 for the privacy budget ε.

The parameter β in Eq.(1) controls the roles MF-MoG

and the sentiment analysis model play in determining

the ﬁnal predicted rating. The lower value of β in-

dicates that the ﬁnal predicted rating is more reliant

on the sentiment analysis model, and the higher value

means it relies more on the MF-MoG model. We eval-

uate the trade-off between β and utility by consider-

ing the value range from 0.1 to 1 for the parameter β.

The utility of a recommendation system is evaluated

by how well it predicts the relevance of an item for a

user. We use F-score as the utility metric to measure

A Local Differential Privacy based Hybrid Recommendation Model with BERT and Matrix Factorization

329

(a) Amazon Instant Video (b) Amazon Toys and Games

Figure 4: BERT-MF-MoG vs MF-MoG RMSE Comparison.

(a) Amazon Instant Video (b) Amazon Toys and Games

Figure 5: BERT-MF-MoG vs MF-MoG F-Score Comparison.

how well the recommendation systems make recom-

mendations that adapt to a user’s choices. Then we

use RMSE to measure the predictive accuracy of our

system.

5.3 Results

In this experiment, we compare the prediction accu-

racy of our proposed recommendation model BERT-

MF-MoG with MF-MoG (Neera et al., 2021) which,

to the best of our knowledge, is the most compara-

ble method as it also uses the exact input perturba-

tion mechanism. We do not compare our method with

other global or local differential privacy based rec-

ommendation systems due to the differences in per-

turbation mechanisms. The privacy budget ε varies

from 0.1 to 3, and we use 0.7 as β values for both

datasets. Fig. 4 and 4 shows the RMSE values for

BERT-MF-MoG (β = 0.7), MF-MoG and the baseline

method, Non-Private BERT-MF. The baseline method

does not perturb the user’s original ratings, and the

BERT review pre-processing takes place at the service

provider’s side as no rating perturbation is performed,

there is no need to combine MoG with MF. As ex-

pected, the prediction accuracy of the two privacy-

preserving methods increases when the privacy bud-

get ε increases. The results also show that BERT-

MF-MoG produces a higher increase in recommen-

dation accuracy for both datasets for all the values

of ε than MF-MoG. This is because BERT-MF-MoG

combines the sentiment classiﬁcation with the MF-

MoG model and predicted rating module takes into

account both ratings and reviews. Fig.5 shows the F-

score values for BERT-MF-MoG and MF-MoG mod-

els. The F-score value increases when the privacy

budget ε increases for both methods. These F-score

values demonstrate again that the BERT-MF-MoG

model provides more accurate recommendations than

MF-MoG for all privacy budget values ε.

Fig.6 and Fig.7 shows the RMSE and the F-score

values respectively when varying β while keeping pri-

vacy budget at ε = 1.5. These ﬁgures show that op-

timal RMSE and F-score values are obtained when

β = 0.7. When β = 0, only the output of the senti-

ment analysis model, P

Sent

, is used to determine the

ﬁnal predicted rating as indicated in Eq.(1), and the

output of the MF-MoG model is not used. Only a

few ratings are used as sentiment labels for the CNN

SECRYPT 2022 - 19th International Conference on Security and Cryptography

330

(a) Amazon Instant Video (b) Amazon Toys and Games

Figure 6: β-RMSE Comparison.

(a) Amazon Instant Video (b) Amazon Toys and Games

Figure 7: β-F-Score Comparison.

model, and they are perturbed at the user side. When

β increases from 0, the contribution of the MF-MoG

model, which uses a large proportion of ratings, starts

to be included in the ﬁnal predicted rating. Hence, a

decrease in RMSE (increase in F-Score) can be ob-

served as β increases in Fig.6 and Fig.7.

When β increases from 0.7 to 1, the RMSE value

increases. This is because the prediction relies more

heavily on the MF-MoG model than it should in this

range. The signiﬁcance of sentiment analysis is un-

derestimated. When β = 1, Eq.(1) uses the P

the ﬁnal predicted rating. The output of the sentiment

classiﬁcation model P

Sent

does not play any role in the

prediction. The same trend can be observed in Fig.7

for the F-score value. The larger the β is, the more

contribution the MF-MoG model makes, and the more

accurate the prediction is until β reaches 0.7, where

the highest F-score and lowest RMSE value are ob-

tained. We use F-score as a utility metric and RMSE

as the accuracy metric for our recommendation sys-

tem. The F-score and RMSE results show that the

BERT-MF-MoG model provides the most accurate

recommendations when β = 0.7 for the two datasets.

The optimal value of Beta should be obtained through

numerical evaluation for different datasets.

6 CONCLUSION

In this paper, we proposed an LDP-based recommen-

dation system that incorporates a deep-learning senti-

ment analysis model into a collaborative ﬁltering al-

gorithm. The system protects the privacy of users and

at the same time offers substantial utility to the ser-

vice provider. The recommendation accuracy of our

proposed model is improved by taking advantage of

sentiment analysis performed on user reviews. The

experiments conducted with two amazon review and

rating datasets demonstrated that the utility of our

proposed recommendation system outperforms that of

the recommendation system based just on ratings for

all the values of privacy budget ε. In future work, we

plan to explore using other techniques such as LSTM

(Long Short Term Memory Networks) in combination

with CNN for sentiment classiﬁcation to further im-

prove the recommendation accuracy.

REFERENCES

Cheng, Z., Ding, Y., Zhu, L., and Kankanhalli, M. (2018).

Aspect-aware latent factor model: Rating prediction

A Local Differential Privacy based Hybrid Recommendation Model with BERT and Matrix Factorization

331

with ratings and reviews. In Proceedings of the 2018

world wide web conference, pages 639–648.

Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K.

(2018). Bert: Pre-training of deep bidirectional trans-

formers for language understanding. arXiv preprint

arXiv:1810.04805.

Dwork, C., Roth, A., et al. (2014). The algorithmic foun-

dations of differential privacy. Found. Trends Theor.

Comput. Sci., 9(3-4):211–407.

Goularas, D. and Kamis, S. (2019). Evaluation of deep

learning techniques in sentiment analysis from twit-

ter data. In 2019 International Conference on Deep

Learning and Machine Learning in Emerging Appli-

cations (Deep-ML), pages 12–17. IEEE.

Holohan, N., Antonatos, S., Braghin, S., and

Mac Aonghusa, P. (2018). The bounded laplace

mechanism in differential privacy. arXiv preprint

arXiv:1808.10410.

Lan, Z., Chen, M., Goodman, S., Gimpel, K., Sharma,

P., and Soricut, R. (2019). Albert: A lite bert for

self-supervised learning of language representations.

arXiv preprint arXiv:1909.11942.

Li, X.-B. and Sarkar, S. (2011). Protecting privacy against

record linkage disclosure: A bounded swapping ap-

proach for numeric data. Information Systems Re-

search, 22(4):774–789.

Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D.,

Levy, O., Lewis, M., Zettlemoyer, L., and Stoyanov,

V. (2019). Roberta: A robustly optimized bert pre-

training approach. arXiv preprint arXiv:1907.11692.

Liu, Z., Wang, Y.-X., and Smola, A. (2015). Fast differen-

tially private matrix factorization. In Proceedings of

the 9th ACM Conference on Recommender Systems,

pages 171–178.

McSherry, F. and Mironov, I. (2009). Differentially private

recommender systems: Building privacy into the net-

ﬂix prize contenders. In Proceedings of the 15th ACM

SIGKDD international conference on Knowledge dis-

covery and data mining, pages 627–636.

Meng, X., Wang, S., Shu, K., Li, J., Chen, B., Liu, H., and

Zhang, Y. (2018). Personalized privacy-preserving so-

cial recommendation. In Thirty-Second AAAI Confer-

ence on Artiﬁcial Intelligence.

Mobasher, B., Burke, R., Bhaumik, R., and Williams,

C. (2007). Toward trustworthy recommender sys-

tems: An analysis of attack models and algorithm ro-

bustness. ACM Transactions on Internet Technology

(TOIT), 7(4):23–es.

Neera, J., Chen, X., Aslam, N., Wang, K., and Shu, Z.

(2021). Private and utility enhanced recommendations

with local differential privacy and gaussian mixture

model. arXiv preprint arXiv:2102.13453.

Osman, N., Noah, S., and Darwich, M. (2019). Contextual

sentiment based recommender system to provide rec-

ommendation in the electronic products domain. In-

ternational Journal of Machine Learning and Com-

puting, 9(4):425–431.

Pal, A., Parhi, P., and Aggarwal, M. (2017). An im-

proved content based collaborative ﬁltering algorithm

for movie recommendations. In 2017 tenth interna-

tional conference on contemporary computing (IC3),

pages 1–3. IEEE.

Paterek, A. (2007). Improving regularized singular value

decomposition for collaborative ﬁltering. In Proceed-

ings of KDD cup and workshop, volume 2007, pages

5–8.

Raghavan, S., Gunasekar, S., and Ghosh, J. (2012). Review

quality aware collaborative ﬁltering. In Proceedings of

the sixth ACM conference on Recommender systems,

pages 123–130.

Salinca, A. (2017). Convolutional neural networks for

sentiment classiﬁcation on business reviews. arXiv

preprint arXiv:1710.05978.

Shen, Y. and Jin, H. (2014). Privacy-preserving person-

alized recommendation: An instance-based approach

via differential privacy. In 2014 IEEE International

Conference on Data Mining, pages 540–549. IEEE.

Shin, H., Kim, S., Shin, J., and Xiao, X. (2018). Pri-

vacy enhanced matrix factorization for recommenda-

tion with local differential privacy. IEEE Transactions

on Knowledge and Data Engineering, 30(9):1770–

1782.

Sutanto, J., Palme, E., Tan, C.-H., and Phang, C. W.

(2013). Addressing the personalization-privacy para-

dox: An empirical assessment from a ﬁeld experiment

on smartphone users. MIS quarterly, pages 1141–

1164.

Wang, C., Duan, Q., Tong, C. H., Di, Z., and Gong, W.

(2016). A gui platform for uncertainty quantiﬁcation

of complex dynamical models. Environmental Mod-

elling & Software, 76:1–12.

Wang, Y., Wang, M., and Xu, W. (2018). A sentiment-

enhanced hybrid recommender system for movie rec-

ommendation: a big data analytics framework. Wire-

less Communications and Mobile Computing, 2018.

Zarzour, H., Al-Ayyoub, M., Jararweh, Y., et al. (2021).

Sentiment analysis based on deep learning methods

for explainable recommendations with reviews. In

2021 12th International Conference on Information

and Communication Systems (ICICS), pages 452–456.

IEEE.

SECRYPT 2022 - 19th International Conference on Security and Cryptography

332