Received 31 January 2023, accepted 12 February 2023, date of publication 16 February 2023, date of current version 23 February 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3246060

Recommendation System Based on Deep

Sentiment Analysis and Matrix Factorization
NING LIU 1,2 AND JIANHUA ZHAO 2
1 Collegeof Economics Management, Shangluo University, Shangluo 726000, China
2 Engineering Research Center of Qinling Health Welfare Big Data, Universities of Shaanxi Province, Shangluo 726000, China

Corresponding author: Ning Liu ([email protected])

This work was supported in part by the Key Research and Development Program of Shaanxi Province under Program 2022GY-073; in part
by the Science and Technology Planning Project of Shangluo City under Grant 2022-J-0009; and in part by the Shangluo Universities Key
Disciplines Project, Discipline name: mathematics.

ABSTRACT In order to solve the problem of data sparsity and credibility in collaborative filtering,
a recommendation system based on sentiment analysis and matrix factorization (SAMF) is proposed in
this paper, which uses topic model and deep learning technology to fully mine the implicit information
in reviews to improve the rating matrix and assist recommendation. Firstly, user topic distribution and item
topic distribution are generated from reviews(consisting user reviews and item reviews) through LDA(Latent
Dirichlet Allocation). The user feature matrix and item feature matrix are created based on topic probability.
Secondly, user feature matrix and item feature matrix are integrated to create user-item preference matrix.
Thirdly, the user-item preference matrix and the original rating matrix are integrated to create the user-item
rating matrix. Fourthly, BERT(Bidirectional Encoder Representation from Transformers) is used to quantify
the sentiment information contained in the reviews and integrate the sentiment information with the user-
item rating matrix, to modify and update the user-item rating matrix. Finally, the updated user-item
rating matrix is used to achieve rating prediction and Top-N recommendation. Experiments on Amazon
datasets demonstrates that the proposed SAMF has better recommendation performance than other classical
algorithms.

INDEX TERMS Reviews, sentiment analysis, matrix factorization, recommendation system.

I. INTRODUCTION limited to text resource recommendation, and it is difficult

With the explosive increase of Internet information, the to mine users’ implicit interests [6], [7].
problem of information overloaded has become increasingly The core idea of collaborative filtering is that users with
severe. To deal with overloading information, recommenda- similar tastes share similar rating distributions toward the
tion system(RS) is produced [1]. RS is a technique that seeks same items [8]. Because it won’t bother about the content of
to predict the rating or preference for a user, and it is used to the items during recommendations process, it can complete
provide suggestions or items to users by exploiting various complex recommendations. Collaborative filtering makes
strategies [2]. significant success and plays a vital role in many applications.
Traditional RS mainly includes content-based RS, collab- The main problem in collaborative filtering is data sparsity
orative filtering RS and hybrid RS [3], [4], [5]. The content- and the cold start issue. Without complete information, it is
based RS provides recommendations that depend on the user hard to recommend efficiently [9], [10]. A sparsity problem
profile and the similarity of the item description. However, arises due to user interactions with a small portion of items in
this method requires effective feature extraction, which is the particular application domain, and the cold start issue is
difficult to finish. At the same time, this method is only due to the lack of data about new entities, i.e., a new item/new
user.
A hybrid RS is an integration of collaborative filtering
The associate editor coordinating the review of this manuscript and and content-based method. The hybrid RS can solve the
approving it for publication was Yiqi Liu . cold start problem of items. While taking into account the
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
16994 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ VOLUME 11, 2023
N. Liu, J. Zhao: Recommendation System Based on Deep Sentiment Analysis and Matrix Factorization

preference relationship between users, it can build a better matrix is constructed to assist recommendation. It can solve
recommendation model. However, traditional RS relies on the sparsity problem of the rating matrix and reduce the error
a large number of interactions between users and items to of rating prediction.
achieve recommendations, such as user purchase records or (2) Bert is used to mine the semantic information contained
user rating records, the results of recommendations depend in the review texts. The user sentiment information is deeply
on the user item interaction matrix. Therefore, the sparsity of integrated into the user-item rating matrix, and the user-item
data and the cold start problem of new users pose challenges rating matrix is updated to improve the credibility of the
to the accuracy and interpretability of recommendation rating matrix and the accuracy of recommendations.
results [11], [12]. The rest of this paper is arranged as follows. In Section II,
In recent years, with the rapid development of deep we discuss the related works. Section III explains the
learning in AI(artificial intelligence) application field, deep methodology used in this study. In Section IV, we present
learning has also become an important technology in the field experiment and simulation results, as well as discussion and
of RS. Compared with the traditional RS, deep learning RS influence of research results. Finally, Section V concludes the
can better mine the latent features of data, obtain deep level work and a suggestion is given for future extension of this
feature description of users and items. The deep learning RS work.
mainly uses some deep learning technologies, such as Auto-
Encoders [13], Restricted Boltzmann Machine(RBM) [14], II. RELATED TECHNOLOGY
Convolutional Neural Networks (CNN) [15] and Recurrent A. LDA
Neural Network (RNN) [16] to build the recommendation LDA is a commonly optimization model for document
model. However, they also face some limitations [17]. topic extraction, which belongs to an unsupervised learning
(1) They all use word2vec or glove to pretreat word vectors. algorithm. It is mainly applied to text topic recognition, text
However, this word vector belongs to a kind of static coding, classification, text similarity calculation and other aspects
the same word is the same parameter expression in different in the field of text mining. By decomposing a collection of
contexts. It leads to deviation in the model’s understanding documents into multiple topics in the form of probability
of semantics. (2) They only use review data as input, and distribution,it performs topic clustering or text classification
fail to fully explore the internal relationship between reviews optimization according to the topic distribution.
data and rating data. At the same time, the interpretability Formally, the following terms are defined [18]:
and scalability of the recommendation results are still the A corpus is a collection of M documents denoted by D =
shortcomings of the deep learning RS. {di |i ∈ {1, 2, . . . , M }}, where di is the ith document consisting
In order to solve the problem of data sparsity and credibility of Ni words. Each word corresponds to a latent topic and the
in collaborative filtering, a recommendation system based topic set Z = {zi |i ∈ {1, 2, . . . , M }}, where zi is the topic set
on sentiment analysis and matrix factorization(SAMF) is corresponding to di . It can be seen
proposed in this paper. SAMF uses topic model and deep P that the total number of
topics in document set D is l = M i=1 count(zi ), and the total
learning technology to fully mine the implicit information number of words is N = M
P
i=1 i . The joint distribution of
N
in reviews and improve the rating matrix. Firstly, user topic all variables in the LDA model is defined as
distribution and item topic distribution are generated from
reviews(consisting user reviews and item reviews) through YNi
P(wi , zi , θi , φ|α, β) = P(θi |α) · P(zij |θi )
LDA(Latent Dirichlet Allocation). The user feature matrix j=1
and item feature matrix are created based on topic probability. · P(φ|β) · P(wij |ϕzij ), (1)
Secondly, user feature matrix and item feature matrix are
integrated to create user-item preference matrix. Thirdly, where α, β follows a prior Dirichlet distribution. wi denotes
the user-item preference matrix and the original rating the word set in the ith document, θi denotes the ‘‘text-
matrix are integrated to create the user-item rating matrix. topic’’ distribution probability of the ith document, and φ
Fourthly, BERT(Bidirectional Encoder Representation from denotes the ‘‘topic-word’’ distribution matrix. P(θi |α) is the
Transformers) is used to quantify the sentiment informa- ‘‘text-topic’’ distribution probability of the ith document
tion contained in the reviews and integrate the sentiment generating from Dirichlet’s prior parameter α.P(zij |θi ) is the
information with the user-item rating matrix, to modify topic probability generated in the jth word of the ith document,
and update the user-item rating matrix. Then the updated depending on the topic distribution θi . P(φ|β) is the ‘‘topic-
user-item rating matrix is used to achieve rating prediction word’’ distribution matrix ϕzij of the topic zij generating
and recommendation. Finally, experiments are conducted on from the Dirichlet distribution parameter β.P(wij |ϕzij ) is the
Amazon dataset to verify the effectiveness of the proposed probability corresponding to the word wij generated from the
algorithm. distribution ϕzij .
The main contributions of this work are listed as follows: The main computational problem of LDA topic model is to
(1) The review texts are used to mine the implicit estimate the latent parameters θ, φ and z by Gibbs sampling
information of users and items, the user-item preference algorithm and variational Bayesian method.

VOLUME 11, 2023 16995

 N. Liu, J. Zhao: Recommendation System Based on Deep Sentiment Analysis and Matrix Factorization

C. BERT
BERT is a pre-trained language representation model and it
can directly obtain the global information of the text through
the modeling of the self attention mechanism. Because it has
no forgetting gate mechanism, all the word information can
FIGURE 1. Schematic diagram of matrix factorization.
be retained. Therefore, BERT can better express the complete
semantic information of the sentence, and also directly find
B. MATRIX FACTORIZATION the correlation features between words from the global word
Among many collaborative filtering recommendation algo- features [22].
rithms, matrix factorization has become the most popular BERT is composed of multiple Transformer layers.
one due to its scalability and easy implementation. It is It conducts multi-level linear transformation on the input
based on the assumption that user preferences are affected vectors to obtain different linear values, then inputs them to
by a small number of latent factors, and the user-item rating the attention module to calculate the attention weight. Finally,
depends on how each characteristic factor is applied to user the output value of the multi-head attention mechanism is
preferences [19], [20]. Matrix factorization can map the user- combined to make another linear change. Any vector input
item rating matrix into two or more low dimensional matrices to the Transformer is processed and output, and Trans (·)
to reduce the dimensions. It can use low dimensional spatial represents all operation procedures in multiple Transformer
data to study the properties of high-dimensional data, mainly layers, which is defined as
including non negative matrix factorization (NMF) [21],
generalized matrix factorization (GMF) and probabilistic Vt = Trans(Wt Xa + bt ), (4)
matrix factorization (PMF). where Xa denotes the input vector, Vt denotes the output
Suppose the RS contains n users and m items. Denote vector, wt denotes the weight, and bt denotes the offset.
R as the rating matrix with n rows and m columns, matrix BERT is composed of multiple transformers stacked
factorization is to decompose the rating matrix Rn×m into together. Bert represents the calculation process in Bert, and
two matrices Un×k and Vk×m (k denotes the dimension of the it is defined as
latent factors and it is much smaller than m and n ). Such that
UV T most closely reconstructs the existing ratings of R and Vb = Bert(Wb Xb + bb ), (5)
performs well in predicting ratings for non-rated items of R.
where Vb denotes the output value of BERT, Xb denotes the
The schematic diagram is shown in FIGURE 1.
input vector, wt denotes the weight, and bt denotes the offset.
When using matrix factorization to model learners, we usu-
ally first construct a matrix with a value of 1/0 based on
III. OUR METHOD
the interactive data, and then decompose the matrix into
A. ALGORITHM FRAMEWORK
two low dimensional matrices. Any application scenario with
interaction or rating behavior can consider using matrix The flow chart of the algorithm SAMF proposed in this
decomposition method. paper is shown in FIGURE 2. SAMF mainly includes the
By learning UiT (1 × k latent-factor vector for user i) and following modules. (1) User topic distribution and item topic
Vj (k × 1 latent-factor vector for item j), the inner product distribution are generated from reviews through LDA to
R̂ij of Ui and Vj is approximated to the actual rating value create user feature matrix and item feature matrix based on
Rij assigned by user i for item j. The calculation method is topic probability. User feature matrix and item feature matrix
defined as are integrated to create user-item preference matrix. (2) The
user-item preference matrix and the original rating matrix are
R̂ij = UiT Vj , (2) integrated to create the user-item rating matrix. (3) Bert is
used to quantify the sentiment information contained in the
where UiT denotes the ith row of U T , Vj denotes the jth column reviews, integrate the sentiment information with the user-
of V , and R̂ij denotes the predicted rating of user i for item j. item rating matrix, and modify and update the user-item
In the process of learning, U and V are calculated by rating matrix. (4) The updated user-item rating matrix is used
minimizing the regularized mean squared error as follows: to achieve rating prediction and Top-N recommendation.
1 X
min L(U , V ) = (Rij − UiT Vj )2 B. IMPLEMENTATION PROCESS
U, V 2M
(i,j)∈O
1) GENERATION OF USER FEATURE MATRIX AND ITEM
1 FEATURE MATRIX
+ λ(∥Ui ∥2 + ∥Vi ∥2 ), (3)
2 User reviews and item reviews are preprocessed. The
where Rij is the rating value assigned by user i for item j, preprocessed review text set is trained by topic model using
O is the set of user–item pairs whose rating value has already LDA, and the number of topics is selected according to
been generated in R, M = |O| is the number of elements in O, the requirements for algorithm execution optimization. The
and λ > 0 is the regularization parameter. topics are represented by the characteristics of the review

16996 VOLUME 11, 2023

N. Liu, J. Zhao: Recommendation System Based on Deep Sentiment Analysis and Matrix Factorization

review space. In this study, the original rating matrix and the
user-item preference matrix are factorized, and the common
user latent factor matrix U is used. That is, the user-item
preference matrix is decomposed into U T P, and the original
rating matrix is decomposed into U T V , where U denotes
the user latent factor matrix, P denotes the preference latent
factor matrix, and V denotes the item latent factor matrix. The
loss function is shown in Equation (11), and the minimum
error is calculated.
|F| |G|
1 XX
L (U , V , P) =
‘
(Rij − UiT Vj )2
2
i=1 j=1

FIGURE 2. Algorithm Flow Chart. |F| |G|

1 XX
+ (Sil − UiT Pl )2
2
i=1 l=1
text with a certain probability, the user topic feature matrix
1 1 1
and the item topic feature matrix are created. The probability + ∥U ∥2 + ∥V ∥2 + ∥P∥2 , (11)
distribution of the ith user review on k topics is define as 2 2 2
(u ) (u ) (u ) where Rij denotes the actual rating matrix.
θ(D(ui ) ) = [θ1 i , θ2 i , . . . , θk i ], (6) Then, the gradient descent method is used to solve
where ui denotes the ith user, D(ui ) denotes the review of the Equation (11), the optimized value of U , V, and P are U ∗ , V ∗
ith user, k denotes the number of topics, F denotes the number and P∗ , respectively. The predicted rating matrix is calculated
of users, and θ denotes probability distribution. according to Equation (12).
For the convenience of expression, Equation (6) is
R∗ = U ∗ (V ∗ )T , (12)
transformed into Equation (7). At the same time, the user
feature matrix is created, as shown in Equation (8), which where R∗ denotes the user-item rating matrix.
denotes the probability distribution of |F| users on k topics.
(u ) (u ) (u ) 4) UPDATING OF USER-ITEM RATING MATRIX INTEGRATING
H (ui ) = [H1 i , H2 i , . . . , Hk i ]. (7)
SENTIMENT ANALYSIS TECHNOLOGY
H = [H (u1 )
,H (u2 )
,...,H ] .
(u|F| ) T
(8)
The user-item rating matrix obtained by Equation (12)
In the same way, the item feature matrix is define as reduces the sparsity of the rating table to a certain extent
through user reviews, but there are still limitations, such
I = [I (v1 ) , I (v2 ) , . . . , I (v|G| ) ]T , (9) as low credibility. Based on the above problems, the
where |G| denotes the number of items. sentiment information contained in user reviews is mined
here. Specifically, the sentiment information contained in
2) GENERATION OF USER-ITEM PREFERENCE MATRIX the reviews is quantified, and integrated with the user-item
The user feature matrix and item feature matrix are integrated rating matrix. The user-item rating matrix is modified and
on K topics, to generate the user-item preference matrix updated, so that the rating matrix can more accurately and
according to Equation (10). Here, S can effectively use the comprehensively represent the truest sentimental tendencies
auxiliary information of user reviews to alleviate the sparsity of users. The specific calculation process is shown below.
of the original rating matrix. First, the user’s review text is input into BERT for word
vector extraction to obtain the word vector representation
S = HI T ∈ R|F|×|G| , (10) of the review text. In each review, each word is converted
into a word segmentation vector EToken_EMB , segment vector
where H denotes the user feature matrix, I denotes the item
ESegment_EMB and position vector EPosition_emb . For ith review,
feature matrix, and S denotes the user-item preference matrix.
three vectors are combined according to Equation (13).
The input review sequence is converted into vector encod-
3) INTEGRATION OF USER-ITEM PREFERENCE MATRIX AND
ing according to Equation (14). After vector encoding is
RATING MATRIX
processed by multi-level transformer transformation, output
The user-item preference matrix and the original rating matrix
word vector from BERT is obtained, as show in Equa-
come from different data sources. One is the probability
tion (15).
distribution, and the other is the rating of users for items.
Here, the integration of the two is achieved. Since the ratings Ej = EToken_emb + Esegment_emb + EPosition_emb . (13)
and reviews have the same user preferences, the user-item
Einput = {E0 , E1 , . . . Ek }. (14)
preference matrix and the original rating matrix share a
common latent factor matrix in the rating space and the Toutput = Trans(Einput ). (15)

VOLUME 11, 2023 16997

 N. Liu, J. Zhao: Recommendation System Based on Deep Sentiment Analysis and Matrix Factorization

TABLE 1. The experimental datasets. Amazon food dataset contains user reviews and user ratings
about food and drinks sold on Amazon.com. The Amazon
Clothing dataset consists of user reviews and user ratings for
clothes, shoes and jewelry product on Amazon.com.
In order to better evaluate the performance of the proposed
algorithm, user reviews and item reviews are preprocessed by
word segmentation and stop words removal. The frequency of
the words are also calculated. Reviews with words less than
3 will be deleted from the dataset, and words with frequencies
Second, the word vector is transferred to BiRNN layer less than 5 will be removed as noise.
to extract the sentiment features of user reviews, and the
sentiment rating is obtained through the Softmax classifier. 2) EXPERIMENTAL SETTING AND EVALUATION INDEX
Given that the input vector of BiRNN is {T0 , T1 , . . . , Tk }, During the learning process, the dataset is randomly divided
the corresponding output sentiment feature vector is {H0 , into training set and test set at a ratio of 7:3, and the average
H1 , . . . , Hk }. The user’s sentiment feature is numerically value is taken as the experiment result after five experiments.
processed to obtain the sentiment rating of the review text, In this study, MAE (Mean Absolute Error) is chosen to test
the computing method is defined as the accuracy of the proposed model and other comparative
models in rating and forecasting. MAE is the average of the
p(si |H , ws , bs ) = soft max(ws H + bs ), (16) absolute difference between the actual value and the predicted
where ws denotes the weight, bs denotes the offset in value in the dataset, and it measures the average of the
sentiment calculation, and si denotes the sentiment rating residuals in the dataset. The calculation method of MAE is
after calculation. shown in Equation (18). The lower values of MAE indicate
Third, the sentiment rating is combined with the user- better performance of models in the accuracy.
item rating to achieve the correction and update of the m,n
1 X
user-item rating matrix. The updating method is shown MAE = Rij − R̂ij , (18)
m
in Equation (17). The sentiment rating si is calculated by i=1,j=1
Equation (16), it is weighted and summed with the rating ri where m denotes the number of users, n denotes the number
from user-item rating matrix. of items, Rij denotes the actual rating value of user i for item j,
riι = (1 − α)ri + αsi , (17) and R̂ij denotes the predicted rating of user i for item j.
In order to verify the effect of the proposed algorithm in
where ri denotes the value from R∗ in Equation (12), α recommendation, F1-score is selected to test user tendencies
denotes the balance factor that measures the weight between and the recommendation accuracy. The F-score measure is
the two points, and rit denotes the user-item rating integrating presented for evaluation as a harmonic mean of retrieval
sentiment factors. and accuracy, and higher values of F1-Score indicate better
Here, the new rating matrix Rnew integrating sentiment performance of models in recommendation.
factors is named updated user-item rating matrix. Precision is defined as Equation (19), Recall is defined as
Equation (20), and F-score is defined as Equation (21), where
5) RATING FORECAST R (U ) indicates N items recommended to user u and T (U )
The updated user-item rating matrix is used to complete indicates the related options for user u in test set.
Top-N recommendation. Based on the predicted rating, P
|R(U ) ∩ T (U )|
a rating sequence from high to low for each item is established u
for each user. In the rating sequence, the items that the user Pr ecision = P . (19)
|R(u)|
has purchased are eliminated, and the first N items in the P u
remaining rating sequence are recommended to the user. |R(U ) ∩ T (U )|
u
Recall = P . (20)
IV. EXPERIMENT
|T (u)|
u
A. EXPERIMENTAL DATA, EXPERIMENTAL SETTING, AND 2 ∗ Pr ecision ∗ Recall
EVALUATION INDICATORS F1 − Score = . (21)
Pr ecision + Recall
1) EXPERIMENTAL DATA
In this study, two real datasets are selected to verify the B. COMPARISON MODEL
experimental performance. They are the Amazon food dataset In order to prove the superiority of the proposed algorithm,
and the Amazon Clothing dataset. The detail of dataset is the proposed SAMF is compared with the other three
shown in Table 1. classical recommendation algorithms.
The Amazon food dataset and Amazon Clothing dataset (1) LFM(Latent Factor Model) [23]. LFM is one of
are offered by Kaggle (https://www. kaggle.com). The the topic models, which decomposes the co-occurrence

16998 VOLUME 11, 2023

N. Liu, J. Zhao: Recommendation System Based on Deep Sentiment Analysis and Matrix Factorization

TABLE 2. The MAE on Amazon food dataset.

TABLE 3. The MAE on Amazon clothes dataset.

matrix into user latent factor matrix and item latent factor
matrix. It calculates the user’s click rate or rating for each
item through the inner product of user latent-factor vector
and item latent-factor vector, and finally completes the
recommendation according to the rating matrix.
(2) SVD(Singular Value Decomposition)++ [24]. On the
basis of SVD, several optimizations have been made in
SVD++. In addition to transforming the matrix decompo-
sition problem into an optimization problem and adding a
regularization term to the loss function, SVD++ also takes
into account that the correlation between items evaluated by
users may affect the prediction of rating. It also introduces FIGURE 3. The comparison results of different methods on Amazon food
dataset.
the influence factor among historical items into the prediction
function.
(3) MFFR(matrix factorization fusing reviews) [25]. MFFR, the MAE value of SAMF decreased by 4.6% on
MFFR is a recommendation algorithm based on matrix average.
decomposition. It also integrates the rating matrix and text It can also be seen from Table 3 that the MAE value of
reviews to assist recommendation. SAMF is also lower than the other three comparison algo-
rithms under different k values on Amazon Clothes dataset.
It shows that SAMF has low error rate in rating prediction and
C. EXPERIMENTAL RESULTS AND ANALYSIS
good performance in rating prediction. Compared with LFM,
Under the same experiment environment, SAMF proposed in the MAE value of SAMF decreases by 17% on average;
this paper is compared with LFM, SVD++, and MFFR. Each compared with SVD++, the MAE value of SAMF decreases
algorithm is tested on Amazon food dataset and Amazon by 11% on average; compared with MFFR, the MAE value
cloths dataset respectively. The experiment is divided into of SAMF decreases by 5% on average.
two cases, including rating prediction experiment and Top-N It can be seen from Figure 3 and Figure 4 that, compared
recommendation experiment. with LFM, SVD++ and MFFR, SAMF is also stable.
In the whole change process of K , the performance of
1) RATING PREDICTION EXPERIMENT SAMF is better than the other three algorithms. It shows
Table 2 and Table 3 show the value of MAE in the rating that SAMF can make full use of text review information to
prediction experiment results. FIGURE 3 shows the change of assist recommendation, fill in the rating matrix, and mine the
MAE with the number of recommendations k in the Amazon implicit sentiments of users. It can more truly and accurately
food dataset, and Figure 4 shows the change of MAE with the understand users’ preferences and has a higher performance
number of recommendations k in the Amazon clothes dataset. of rating prediction.
The results are presented as shown in Table 2. Compared
with LFM, SVD++ and MFFR, the MAE value of SAMF 2) TOP-N RECOMMENDED EXPERIMENT
is also lower than the other three comparison algorithms The values of F1-Score are shown in Table 4 and Table 5.
under different k values on Amazon food dataset. It shows FIGURE 5 shows the change of F1-Score with the number
that the proposed algorithm SAMF has low error rate in of recommendations k in the Amazon food dataset, and
rating prediction and good performance in rating prediction. Figure 6 shows the change of F1-Score with the number of
Compared with LFM, the MAE value of SAMF decreases by recommendations k in the Amazon clothes dataset.
15% on average; compared with SVD++, the MAE value It can be calculated from Table 4 that on the Amazon
of SAMF decreases by 14.6% on average; compared with food dataset, the F1-Score value of SAMF is higher than that

VOLUME 11, 2023 16999

 N. Liu, J. Zhao: Recommendation System Based on Deep Sentiment Analysis and Matrix Factorization

TABLE 4. F1-Score on Amazon food dataset.

TABLE 5. F1-Score on Amazon clothes dataset.

FIGURE 4. The comparison results of different methods on Amazon

clothes dataset.
FIGURE 6. The comparison results of different methods on Amazon
clothes dataset.

assist recommendation and fill in the rating matrix, and

it can effectively solve the problem of data sparsity and
improve the recommendation performance. (2)Compared
with MFFR, SAMF can fully mine the implicit information
in reviews, analyze the semantic features of reviews by
using rating, reviews and other resources, and fully mine
the semantic information of reviews. (3)SAMF uses the
sentiment information contained in the reviews to update the
rating table, so that the rating table can accurately reflect
the user’s preferences and recommend items that may be
FIGURE 5. The comparison results of different methods on Amazon food of interest to users. So SAMF has better recommendation
dataset. performance.
It can be seen from Figure 5 and Figure 6 that SAMF
of the other three comparison algorithms under different k is also stable compared with LFM, SVD++ and MFFR.
values. Compared with LFM, the F1-Score value of SAMF The performance of SAMF is better than the other three
has an average increase of 5.8%; compared with SVD++, the algorithms in the whole change process of K . Especially,
F1-Score value of SAMF has an average increase of 8.2%; compared with MFFR, the performance of SAMF is always
compared with MFFR, the F1-Score value of SAMF has an better. It shows that SAMF can make full use of the text
average increase of 5.9%. review information, excavate the implicit sentiments of users,
It can be calculated from Table 5 that on Amazon Clothes understand users’ preferences more truly, and implement
dataset, SAMF is higher than the other three comparison recommendations more accurately.
algorithms under different k values. Compared with LFM, the
F1-Score value of SAMF has an average increase of 18.4%; V. CONCLUSION
compared with SVD++, the F1-Score value of SAMF has In order to further improve the performance of RS and solve
an average increase of 19.2%; compared with MFFR, the the problem of data sparsity and credibility in collaborative
F1-Score value of SAMF has an average increase of 3.5%. filtering, we propose a deep learning RS based on sentiment
There are several reasons. (1)Compared with LFM and analysis and matrix factorization in this study. The topic
SVD++, SAMF makes full use of review text data to model LDA and deep learning technology are used to

17000 VOLUME 11, 2023

N. Liu, J. Zhao: Recommendation System Based on Deep Sentiment Analysis and Matrix Factorization

fully mine the implicit information in text reviews to [16] L. Huang, B. Jiang, S. Lv, Y. Liu, and D. Li, ‘‘Survey on deep
improve the rating matrix and assist in recommendation. learning based recommender systems,’’ Chin. J. Comput., vol. 41, no. 7,
pp. 1619–1647, 2018.
Experimental results show that the proposed algorithm has [17] S. Seo, J. Huang, H. Yang, and Y. Liu, ‘‘Interpretable convolutional neural
better performance. networks with dual local and global attention for review rating prediction,’’
In the future, we intend to extend our work into two main in Proc. 11th ACM Conf. Recommender Syst., Como, Italy, Aug. 2017,
pp. 297–305.
directions. The first direction is to design new methods to [18] D. M. Blei, A. Y. Ng, and M. I. Jordan, ‘‘Latent Dirichlet allocation,’’
use tags and the knowledge graph as auxiliary information to J. Mach. Learn. Res., vol. 3, pp. 993–1022, Mar. 2003.
improve recommendation performance. The other direction [19] D. Lee and H. Seung, ‘‘Algorithms for non-negative matrix factorization,’’
in Proc. Adv. Neural Inf. Process. Syst., Denver, CO, USA, 2000,
is to effectively describe the dynamic characteristics of pp. 556–562.
user preferences in tag recommendation scenarios, and [20] S. Zhang, W. Wang, J. Ford, and F. Makedon, ‘‘Learning from incomplete
propose a tag recommendation model based on dynamic user ratings using non-negative matrix factorization,’’ in Proc. SIAM Int. Conf.
Data Mining, Bethesda, MD, USA, Apr. 2006, pp. 549–553.
preferences. [21] D. Lee and H. Seung, ‘‘Learning the parts of objects by non-negative matrix
factorization,’’ Nature, vol. 401, no. 10, pp. 789–791, 1999.
REFERENCES [22] N. Liu and J. Zhao, ‘‘A BERT-based aspect-level sentiment analysis
algorithm for cross-domain text,’’ Comput. Intell. Neurosci., vol. 2022,
[1] Z. Batmaz, A. Yurekli, A. Bilge, and C. Kaleli, ‘‘A review on deep learning
pp. 1–11, Jun. 2022.
for recommender systems: Challenges and remedies,’’ Artif. Intell. Rev.,
[23] Y. Koren, ‘‘Factorization meets the neighborhood: A multifaceted collab-
vol. 52, no. 1, pp. 1–37, Jun. 2018.
orative filtering model,’’ in Proc. 14th ACM SIGKDD Int. Conf. Knowl.
[2] E. Çano and M. Morisio, ‘‘Hybrid recommender systems: A systematic
Discovery Data Mining, vol. 2008, pp. 426–434.
literature review,’’ Intell. Data Anal., vol. 21, no. 6, pp. 1487–1524,
[24] R. Kumar, B. K. Verma, and S. Sunder Rastogi, ‘‘Social popularity based
Nov. 2017.
SVD++ recommender system,’’ Int. J. Comput. Appl., vol. 87, no. 14,
[3] X. Yu, F. Jiang, J. Du, and D. Gong, ‘‘A cross-domain collaborative filtering
pp. 33–37, Feb. 2014.
algorithm with expanding user and item features via the latent factor space
of auxiliary domains,’’ Pattern Recognit., vol. 94, pp. 96–109, Oct. 2019. [25] H. Wang, Z. Hong, and M. Hong, ‘‘Research on product recommendation
based on matrix factorization models fusing user reviews,’’ Appl. Soft
[4] B. Alhijawi, G. Al-Naymat, N. Obeid, and A. Awajan, ‘‘Novel predictive
Comput., vol. 123, Jul. 2022, Art. no. 108971.
model to improve the accuracy of collaborative filtering recommender
systems,’’ Inf. Syst., vol. 96, Feb. 2021, Art. no. 101670. [26] X. Cai, L. Xie, R. Tian, and Z. Cui, ‘‘Explicable recommendation
based on knowledge graph,’’ Expert Syst. Appl., vol. 200, Aug. 2022,
[5] F. Liang, E. Yang, W. Pan, Q. Yang, and Z. Ming, ‘‘Survey of recommender
Art. no. 117035.
systems based on federated learning,’’ Scientia Sinica Informationis,
vol. 52, no. 5, pp. 713–741, 2022.
[6] Y. Zhang, X. Zhu, C. Xu, and S. Dong, ‘‘Hybrid recommendation
approach based on deep sentiment analysis of user reviews and multi-view
collaborative fusion,’’ Chin. J. Comput., vol. 42, no. 6, pp. 1316–1333,
2019.
[7] Z. Xie, C. Jin, and Y. Liu, ‘‘Personalized knowledge recommendation NING LIU received the master’s degree in com-
model based on constructivist learning theory,’’ J. Comput. Res. Develop.,
puter software and theory from the School of Infor-
vol. 55, no. 1, pp. 125–138, 2018.
mation Science and Technology, Northwestern
[8] X. Zheng, M. Wang, C. Chen, Y. Wang, and Z. Cheng, ‘‘EXPLORE:
EXPLainable item-tag CO-REcommendation,’’ Inf. Sci., vol. 474, University, China, in 2017. She is currently an
pp. 170–186, Feb. 2019. Assistant Professor with the School of Economics
[9] B. Sarwar, G. Karypis, J. Konstan, and J. Riedl, ‘‘Item-based collaborative and Management, Shangluo University, China.
filtering recommendation algorithms,’’ in Proc. 10th Int. Conf. World Wide Her current research interests include natural
Web, Hong Kong, Apr. 2001, pp. 285–295. language processing, recommendation system,
[10] Y. Koren, R. Bell, and C. Volinsky, ‘‘Matrix factorization techniques sentiment analysis, semi-supervised learning, and
for recommender systems,’’ IEEE Comput., vol. 42, no. 8, pp. 30–37, deep learning.
Aug. 2009.
[11] H. Wang, N. Wang, and D.-Y. Yeung, ‘‘Collaborative deep learning
for recommender systems,’’ in Proc. 21st ACM SIGKDD Int. Conf.
Knowl. Discovery Data Mining, Sydney, NSW, Australia, Aug. 2015,
pp. 1235–1244.
[12] K. Miyahara and M. Pazzani, ‘‘Collaborative filtering with the simple
JIANHUA ZHAO received the Ph.D. degree
Bayesian classifier,’’ in Proc. 6th Pacific Rim Int. Conf. Artif. Intell., Berlin,
in computer science and technology from the
Germany, 2000, pp. 679–689.
School of Computer, Northwestern Polytechnical
[13] S. Sedhain, A. Menon, S. Sanne, and L. Xie, ‘‘AutoRec: Autoencoders
meet collaborative filtering,’’ in Proc. 24th Int. Conf. World Wide Web, University, China, in 2014. He is currently a
Florence, Italy, 2015, pp. 111–112. Professor with the Engineering Research Center
[14] A. Biswal, M. Borah, and Z. Hussain, ‘‘Music recommender system using of Qinling Health Welfare Big Data, Universities
restricted Boltzmann machine with implicit feedback,’’ Adv. Comput., of Shaanxi, China. His current research interests
vol. 122, pp. 367–402, 2021. include natural language processing, recommen-
[15] J. Cheng, P.-S. Wang, G. Li, Q.-H. Hu, and H.-Q. Lu, ‘‘Recent advances dation systems, sentiment analysis, deep learning,
in efficient computation of deep convolutional neural networks,’’ Frontiers transfer learning, and semi-supervised learning.
Inf. Technol. Electron. Eng., vol. 19, no. 1, pp. 64–77, Jan. 2018.

VOLUME 11, 2023 17001