Downloaded from [Link]

ir
[Link] [Link]

Multimedia Tools and Applications (2023) 82:8709–8728

[Link]

1209: RECENT ADVANCES ON SOCIAL MEDIA ANALYTICS

AND MULTIMEDIA SYSTEMS: ISSUES AND CHALLENGES

DeepLSGR: Neural collaborative filtering

for recommendation systems in smart community

Srinidhi Hiriyannaiah1 · G M Siddesh1 · Srinivasa K. G.2

Received: 24 December 2020 / Revised: 21 August 2021 / Accepted: 9 September 2021 /

Published online: 3 January 2022
© The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2021

Abstract
In the field of Data science and online world, Recommendation Systems (RS) play an
important role among the various e-commerce applications. Data sparsity often leads to
the problem of precise recommendations in RS as there will be either less number of users
or ratings. Collaborative filtering (CF) is one of the key techniques that are used for the RS
with the pre-requisite of the adequate information of the users and items. Deep Learning
(DL) models haved paved the way for analysis and prediction of sequential textual infor-
mation in various applications. Hence, CF combined with DL approaches are also being
explored to solve the problem of data sparsity in RS with various challenges of analysis
and prediction of the sequential information. This paper considers the problem of data
sparsity with a novel neural CF based DeepLSGR model to provide better recommenda-
tions. It is a bi-directional model composed of stacked hidden layers with Long-short term
memory (LSTM) and Gated recurrent unit (GRU) and provide recommendations based on
the prediction of rating using the textual reviews from the users. It provided an accuracy
of 97%, recall of 61% and RMSE of 0.87 for the experiments conducted on the Amazon
Fine Food Reviews and OpinRank datasets. The results of the comparison with the exist-
ing works evidently demonstrate that the DeepLSGR provides improved recommendations.

Keywords Recommendation systems · Collaborative filtering · Deep learning · LSTM ·

GRU​· Web recommenders

* Srinidhi Hiriyannaiah
[Link]@[Link]
1
M S Ramaiah Institute of Technology, Affiliated to Visvesvaraya Technological University,
Karnataka, Bengaluru, India
2
National Institute of Technical Teachers Training & Research, Chandigarh, India

13
Vol.:(0123456789)
 Downloaded from [Link]
[Link] [Link]

8710 Multimedia Tools and Applications (2023) 82:8709–8728

1 Introduction

In the past few years, e-commerce applications have flourished in wide areas of world mak-
ing it more attractive for the online users. There is a paradigm shift in the retail to e-retail
business since most of the users are tend to shop online [11]. In this context, it is important
for the e-commerce applications to retrieve the data of the users and provide recommenda-
tions. RS in this regard play an important role in the e-commerce domain [9]. A RS helps
in assisting the users to purchase the products/items based on the previous purchase his-
tory of them using learning algorithms [14]. For example, if a user has purchased a ‘Harry
potter’ book, then the user interests lie in the Science fiction and such books are recom-
mended. It should take into account different categories of products for each user as the
preference differ from each other.
Deep learning (DL) has paved the way for many areas in computer vision and Natural
language processing (NLP) [41]. DL architectures such as Convolutional neural networks
(CNN), recurrent neural networks are used in NLP applications for Parts of speech (POS)
tagging [17], sentiment classification [42], question answering [40], speech recognition
[30] and others. DL models are used for extracting the semantic and context level mean-
ings from the words and phrases in sentences and documents. RS depend on the reviews
and ratings given by the users. The different DL models can help in the understanding of
the reviews and the dependencies between them. DL models are being used in these days to
understand the semantics among the sentences in the document rather than the traditional
NLP techniques [30].
RS use techniques such as collaborative filtering, content-based filtering and hybrid fil-
tering to provide recommendations [32]. In Collaborative Filtering (CF), different informa-
tion on the user’s behaviour is collected based on the reviews of the products purchased
from them. The information is usually gathered in terms of ratings that can be on a scale
of 5 or 10 [32]. In content-based filtering, recommendations are provided based on the
description of the products purchased and the user information. It does not use any informa-
tion related to the reviews and ratings by the user [33]. The main drawback of this approach
is it depends on the content of the product and user profile. The scarcity of the information
may lead to ambiguous recommendations [26]. Hybrid based filtering approaches combine
both collaborative and content-based filtering to provide recommendations. It is used to
solve problems in recommendation systems such as cold start, data sparsity faced by the
content based and collaborative based filtering approaches.
A typical issue that arises in the RS is the data sparsity. It refers to the problem where
there is small amount of data for each user as compared to the overall size of the corpus. It
occurs due to lack of information about the users and the items related to the purchase his-
tory of the users and the items. CF is one of the widely used methods for recommendation
systems. In this method, the reviews given by a user to a set of items/products is provided
as the input to predict the ratings of the target item/product that is not rated by the users.
It lacks in improvising the recommendations as it is based on the implicit feedback of the
users.
The motivation of the paper is to improve the precision of recommendations by integrat-
ing the CF with DL models. The main contribution of the proposed system is the Deep-
LSGR hybrid model endowed with two hidden layers with LSTM and GRU for improved
recommendations. The results are demonstrated with the datasets Amazon Fine Food
Reviews and OpinRank that shows the improvement in the recommendations. The paper is
further organized as follows. In Sect. 2 the related work in the area of CF and its problems

13
 Downloaded from [Link]
[Link] [Link]

Multimedia Tools and Applications (2023) 82:8709–8728 8711

with DL approaches are discussed. Section 3 discusses the proposed methodology using
the neural network for CF. In the Sect. 4 the experiment and the results are discussed that
compares the efficiency of the hybrid model with the existing ones. The paper is concluded
with the results of the proposed model.

2 State of art methods

Data Sparsity is one of the key challenges of the RS as there is only small amount of infor-
mation of the users and ratings for the recommendations. The base technique of CF [1, 15,
35] in combination with other methods are usually experimented with different datasets.
Hence, the precision of the recommendations is based on the implicit feedback of the users
and thus lacks in improvising the precision of the recommendations [13, 28]. Topic model-
ling methods such as Latent Dirichlet Allocation (LDA) [4] is used to solve the problem
of data sparsity by integrating customer reviews into CF algorithms [6, 24]. These meth-
ods treat the customer reviews as Bag-of-Words model to infer the ratings. Some of the
attempts have been made with CNNs to learn the profile of the users in addressing the data
sparsity problem in RS [3] by transforming the user profile into a sequence of words. A
descriptive based CNN was used to integrate the product description and ratings for pro-
viding recommendations [19]. The drawbacks of these approaches are the ignorance of the
sequential information present in the customer review that can be used for improving the
precision of the recommendations. The Table 1 shows the comparison of the state of the art
techniques with the proposed DeepLSGR solution.

3 Related work

RS are one of the prominent research areas since collaborative filtering technique was
introduced. There has been much research being carried out in the area of recommendation
systems using new approaches in the current decade. RSs can be classified into three cat-
egories namely collaborative filtering; content-based filtering and hybrid systems. Content
based RS [16] are based on the profile of the users and items. It recommends the items
that are similar to the items liked by the user in the past using the history of transactions
of items purchased by the user. On the other hand, collaborative filtering RS recommends
the items using common preferences of users and the items purchased by them [31, 44].
Hybrid systems use a combination of content based and collaborative filtering methods
for recommendation [44]. In the case of hybrid systems, the characteristics are studied
using the ratings of items, similar users and click streams. In this paper, the proposed sys-
tem for RS uses the hybrid approach for recommendation of items to the users. The pro-
posed system uses the review of the users as the core data for the prediction of ratings and
recommendations.
The main stream of problem in RS is Data Sparsity which is one of the key challenges
in the RS [44]. It deals with the imbalance in the ratings of the users against the total
number of items. It is difficult to recommend the items using few ratings and fewer ratings
will lead to data sparsity problem in RS. One of the ways it can be solved is by using the
information available in the review which is in textual format. The studies [31, 44] suggest
the textual information in the review helps in improving the accuracy of the prediction of
the RS where there are few users and ratings [39]. It has been studied that the other latent

13
 8712

13
Table 1  State of the art techniques
Downloaded from [Link]

Techniques Drawbacks Proposed solution

User-item A user-item rating matrix is needed and hence dependent on the The DeepLSGR approach requires only user reviews for the recom-
collaborative filtering (CF) database of the users and the items of a catalogue. It does not mendations and works well with the data sparsity of the users
[1, 15, 35] provide precise recommendations for less data
Latent Dirichlet allocation (LDA) [4] These methods treat the customer reviews as Bag-of-Words The proposed DeepLSGR uses the DL based approach for the
model to infer the ratings. Hence, it considers only certain prediction of ratings and improved recommendations
keywords for the recommendations
Descriptive CNN [19] There is ignorance of the sequential information present in the The proposed approach of DeepLSGR is bi-directional and consid-
user reviews with the CNN based approach ers the sequential information for the precise recommendations
[Link]

Multimedia Tools and Applications (2023) 82:8709–8728

[Link]
 Downloaded from [Link]
[Link] [Link]

Multimedia Tools and Applications (2023) 82:8709–8728 8713

factors that are embedded in the text help in topic modelling and conversion into ratings.
In [25, 36], the studies have been conducted in exploiting the textual information of the
review into rating that uses the bag-of-words model as the main stream method for the
conversion of review into ratings. It uses a set of units to spot the keywords in the review
text that has high potential to influence the users for RS. The limitations of these methods
lies in the lexical similarity of the review texts. The topic modelling using bag-of-words
approach has limited capacity of feature extraction from the reviews of new users and items
and it does not take into account of the similarity measures for recommendations. The pro-
posed DeepLSGR model in this paper helps in the prediction of ratings based on the tex-
tual information in the reviews from the users.
Recent advances in deep learning based recommender systems have gained significant
attention by overcoming obstacles of conventional models and achieving high recommen-
dation quality [43]. The recent works of DL methods for RS can be categorized based
on the various neural building blocks such as Multi-layer perceptron (MLP) [2], Auto
encoders (AEs) [7], Convolution neural networks (CNNs) [5], Recurrent neural networks
(RNNs) [27], Restricted Boltzmann Machine (RBM) [12]. The data sparsity problems can
be well solved using RNNs for recommendation systems [34, 38]. RNNs have various sub-
types such as LSTM and GRU. GRU based recommendation systems are studied in [37]
where the likelihood of the next user session based on the previous session is calculated. In
[23] experiments were conducted on the RNN based models by modelling the text reviews
and ratings simultaneously using LSTM as the neural network. One of the real-time based
RS based on DL is a hashtag RS in [23]. A LSTM was used to extract the features from a
given image of a tweet and recommend hashtags. In this paper, the proposed system uses a
hybrid RNN with LSTM and GRU as the hidden layers for a RS that is suitable to solve the
problem of data sparsity.
A summary of the related work with the drawbacks for the data sparsity problem in
RS is as shown in the Table 2. The comparison of the proposed system is carried out with
the works [10, 21] which consider Probabilistic matrix factorization (PMF) and Collabora-
tive Topic Regression (CTR) as the methods for RS. The main advantages of the proposed
DeepLSGR neural network for RS are of twofold. First, LSTM neural nets are capable of
generating a sequence of text based on the existing textual review. Second, GRU neural
nets are capable of predicting the likelihood of the next user session based on the click
stream. The proposed integration of DL and CF with the DeepLSGR model is discussed in
the further sections.

4 Proposed CFDL recommendation framework using DeepLSGR

The proposed CFDL system that integrates the CF and DL is described in this section. The
reviews given by the users are extracted into latent factors to estimate the ratings given by
them. It is accomplished using a RNN based model with LSTM and GRU coupled with
each other as the hidden layers in RNN. The model is trained to estimate the ratings with
minimum prediction error and then CF is carried using neighbourhood classification. The
different notations that are used as the values for training the DeepLSGR model are as
shown in the Table 3. These notations are used for the CF and recommendation algorithms
that are explained further. The overall methodology using the DeepLSGR model is as
shown in the Fig. 1 that involves the following steps. Each of the steps with its details are
discussed in the further sections.

13
 8714

13
Table 2  Summary of related work
Downloaded from [Link]

Related work Drawbacks Proposed solution

RBM based modelling recommendation [12, 34] The problem of unobserved ratings with the number of DeepLSGR model learns the complex relationships between
users and items grows linearly while the number of known the users and items that are sparse in nature using the hid-
ratings typically scales linearly with the number of users den layer semantic factors from both LSTM and GRU​
Topic based modelling [10, 23] The main drawback of topic modelling is it uses bag-of- The proposed DeepLSGR model is a RNN based sequence-
words approach for predictions. It has limited capacity of to-sequence prediction rather than the typical Bag-of-words
feature extraction from the reviews of new users and items approach for the predictions
Collaborative based topic modelling [21] The drawback arises on the computation of similar values The similarities are calculated using the predictions of the
between the two reviews to provide recommendations DeepLSGR model with pearson correlation coefficient for
ranking the recommendations
[Link]

Multimedia Tools and Applications (2023) 82:8709–8728

[Link]
 Downloaded from [Link]
[Link] [Link]

Multimedia Tools and Applications (2023) 82:8709–8728 8715

Table 3  Notations used for collaborative filtering using DeepLSGR model

Notation Explanation

{<U1,Rv1>, <U2,Rv2>, <U3,Rv3>, A set of reviews <Rv1, ­Rv2, …, ­Rvn> for the users <U1, ­U2, …, ­Un>
……, <Un,Rvn>}
Eu,r Initial embedding matrix for the users and reviews given by them
using the gLove model
Mui ri The updated rating matrix of the users ­ui and their ratings ­ri obtained
by the DeepLSGR model
Ncr Global matrix of the users and the predicted ratings used for col-
laborative filtering
At Average rating of the target item t
Ar Average rating of the remaining item r for all the co-rated users
Pcc(sim) Pearson correlation co-efficient to obtain the similar items using the

Fig. 1  Proposed CFDL recommendation using DeepLSGR

(1) In the pre-processing, the reviews given by the users are collected and pre-processed
to get the tokens of words necessary for identifying the ratings.
(2) The DeepLSGR neural network now examines the different keywords and converts the
reviews into ratings.
(3) The ratings are used by the collaborative filtering module for recommendations using
the similarity metrics.

13
 Downloaded from [Link]
[Link] [Link]

8716 Multimedia Tools and Applications (2023) 82:8709–8728

4.1 Pre‑processing and embedding

The text reviews contains some of the words that are not useful in conversion of the rat-
ings. For example, consider the review ‘The food was great and tasty yet too high in price’.
Here, the key phrases that are necessary for the conversion are great, tasty and price. The
stopwords such as ‘is’, ‘was’ and others need to be removed and also these reviews need
to be converted into word vectors initially using a embedding process. Word embedding
is process of representation of text that contains the words with same meaning in a similar
way. In a coordinate system, it represents the words that are related to each other. How-
ever, word embedding depends on the corpus chosen for the comparison of similar mean-
ing words. For word-vector initialisation, the gloVe embedding is used [29]. It is an unsu-
pervised learning algorithm that is trained on aggregated global word-word co-occurrence
statistics. It learns the vector representation or the geometrical code for words from their
co-occurrence information i.e. how frequently they are found in a text together. The word
embedding takes the reviews as input sequence of words <w1, ­w2, ­w3, …, ­wT> and maps
each review into its respective k-dimensional vector representation ­xi ∈ ­Ru,k.
The algorithm for initializing the embedding matrix is as shown in the Algorithm 1.
Initially, the input is provided as list of users and the reviews given by them in the form
{<U1,Rv1>, <U2,Rv2>, <U3,Rv3>, …… <Un,Rvn>}. A list of word indexes are formu-
lated from the reviews. A glove vector is used as a dictionary that is used to check for
the word indices from the review. The word indices if present are added to the embed-
ding matrix ­Eu,r. The Algorithm 1 takes each review ­Rvi and convert it into a list of word
indices ­[w1, ­w2, ­w3, .., ­wn]. These word indices are compared with the glove dictionary
to obtain the embedding matrix E ­ u,r. It is the initial Embedding matrix for the users and
reviews given by them using the gLove model. This matrix is taken as the input by the
DeepLSGR model for the prediction of ratings. The DeepLSGR model is explained in the
further sections.

Algorithm 1: CFDL_Embedding
Input: Reviews provided by the users {<U1,Rv1>,<U2,Rv2>,<U3,Rv3>,……<Un,Rvn>}
Output: Ratings for reviews in the test dataset{<U1,r1>,<U2,r2>,<U3,r3>,……<Un,rn>}=
Ru,p
Begin
Read reviews and replace missing values in the training set with a default
parameter
(<Rvn>,[w1, w2, w3,…. wn]) <- Modify reviews into a list of word indexes
{Gv}<- Import the glove vector file as a dictionary
for each word in (<Rvn>,[w1, w2, w3,…. wn]):
Ev<- word in {Gv}
Eu,r ← Eu,r + Ev
Ru,p<- RNN (Eu,r ,h1=GRU, h2=LSTM)
End

4.2 DeepLSGR neural network model

In this section the DeepLSGR model that is used for converting the user reviews into ratings is
discussed. The neural network model DeepLSGR with a combination of LSTM and GRU is as
shown in the Fig. 2. It is implemented using a bi-directional RNN that consists of LSTM and
GRU as the hidden layers in both forward and backward directions. Every layer of the GRU
treats the memory sequence of the previous LSTM layer as the input sequence and computes

13
 Downloaded from [Link]
[Link] [Link]

Multimedia Tools and Applications (2023) 82:8709–8728 8717

Fig. 2  DeepLSGR model with stacked LSTM and GRU layers

its own memory representation. The forward DeepLSGR reads the input sequence in the as in
the usual order, whereas the backward DeepLSGR reads it in the reversed order. The activa-
tions of the DeepLSGR in forward and backward directions are denoted as h��⃗t , ⃖�� ht respectively
for the time steps t. In the forward DeepLSGR, the model calculates the sequence of the for-
ward hidden states by arranging the input sequence from first word to the last word as shown
in the Eq. 1. In the backward phase, the DeepLSGR model takes the input sequence in reverse
order and computes the backward hidden states as shown in the Eq. 2. For the final prediction,
the average output at is computed from the last layer of DeepLSGR model in both the direc-
tions as shown in the Eq. 3. A prediction module is used for the final prediction of the ratings.
The prediction module uses the averaged vector at and the transformed annotations st by com-
puting the dot products and assigns a weighting score using the softmax function as shown in
the Eqs. 4 and 5 where U ∈ ­Eh. For each user and review {<U1,Rv1>, <U2,Rv2>, <U3,Rv3>,
…… <Un,Rvn>}, the final predicted ratings from the equation are concatenated into Mui ri
which is the updated rating matrix of the users u­ i and their ratings r­i obtained by the Deep-
LSGR model for the collaborative filtering to provide further recommendations.
( )
����⃗ �����(i)⃗h������
������(i)⃗ h(i−1) + V (i)⃗ ����(i)⃗
h(i)
t = f W t t−1
+ b (1)

( )
⃖���� ⃖������ (i) ⃖������
⃖����� ⃖����
h(i)
t = f W (i) (i−1)
ht + V h (i)
t+1
+ b (i)
(2)

13
 Downloaded from [Link]
[Link] [Link]

8718 Multimedia Tools and Applications (2023) 82:8709–8728

( )
⃗ ⃖��������
��������
(last) (last)
at = average ht , ht (3)

ak .sk
pkt = ∑T t t (4)
k k
t=1 at .st

y = softmax UpT + c (5)

( )

4.3 Collaborative filtering

The rating matrix that is obtained from the DeepLSGR model is used for collaborative fil-
tering further using the Algorithm 2. For each of the users <U1,U2, ­U3, …, ­Un> and prod-
ucts <P1, ­P2, ­P3, .., P
­ n> the matrix of the ratings of the users ­Rui and the matrix of the
ratings of the items R ­ pj are joined to obtain the global rating matrix N ­ cr. In order to obtain
the matrix ­Ncr the values are normalized using the differences of each rating r­ i and r­ j with
the sum of all values in R ­ u,p. The pearson correlation co-efficient is calculated between the
di and dj to calculate the similarity further using Eq. 6 where ­Rjt is the rating of the target
product t by user i, R­ jr is the rating of the remaining item r by user i, A ­ t is the average rating
of the target item t for all the co-rated users, A ­ r is the average rating of the remaining item
r for all the co-rated users, and m is the number of all rating users to the items i and r. The
common support for each pair of items is the number of users who have reviewed both.
This value is used to modify the similarity between the items. If the number of common
reviewers of a item is low, then the similarity of the target item could be artificially high
solely based on the common reviewer. The high similarity needs to be reduced. For this
reason, the Pearson co-efficient (similarities) are shrunk using a regularization parameter λ.
The similarities Sj for m number of users are calculated for each product ­pj are then regu-
larized using the Eq. 7 where N ­ cr is the number of common reviewers (common support),
Pcc (sim) is the Pearson co-relation coefficient and λ is the regularization parameter.
∑m � �� �
j=1 Rjt − At Rjr − Ar
Pcc (sim) = � (6)
∑m � �2 � �2
j=1
Rjt − At Rjr − Ar

∑m Ncr ∗ Pcc (sim)

Sj = (7)
j=1 Ncr + 𝜆

5 Experiment and results

5.1 Dataset and experimental setting

The experiments were conducted on the Amazon Fine Food Reviews dataset [22]
and OpinRank Dataset [8] as shown in the Table 4. The Amazon dataset includes

13
 Downloaded from [Link]
[Link] [Link]

Multimedia Tools and Applications (2023) 82:8709–8728 8719

Table 4  Statistics of the datasets Dataset #users #reviews %sparse

Amazon Fine Food 256,059 568,454 27%

Reviews
OpinRank 356,059 301,230 37%

information about the product and the user, ratings, and review in textual form. It
contains 568,454 reviews of fine foods for 256,059 users on 74,258 products for over
a decade up to October 2012. The columns included are row id, product id, user id,
helpfulness factors, time, ratings and full text reviews. The OpinRank dataset contains
full reviews for cars and hotels collected from Tripadvisor (~ 259,000 reviews) and
Edmunds (~ 42,230 reviews). The dataset was divided for training and testing purposes
in the ratio of 80:20 and the same has been used for all models under consideration.

Algorithm 2: CFDL_Recommendation
Input: Ru,p (<U1,U2, U3.., Un>,<P1, P2, P3.., Pn>)
Output: Recommended Product IDs and Product Names
For each user ui in the Ru,r:
For each product pj in the Ru,p:
NCr ← Rui join Rpj
ri<-Ncri
rj<-Ncrj
= −∑

= −∑
PCC[pj] <- −
endfor
endfor
For all products in Ru,p
Find all the similarities for Product ID
∗ ( )
Sp <- ∑ =1 =
+
endfor
For all users in Ru,p
Su,k <-Find the k number of highest similarities using k_nearest_neighbour()
Ru,p <- Recommend k-items from Su,k
endfor

The text reviews has to be converted into a word vector for the rating conversion
purposes. For the experiments, a gloVe embedding is used for the word-vector ini-
tialization. It is used as an unsupervised learning algorithm for training the aggregated
global word-to-word co-occurrence. This co-occurrence information is used to learn
the vector representation and the geometrical code for the words. The advantage of
using the gloVe embedding is it gives the co-occurrence matrix of the words that are
frequently used. In textual reviews, gloVe embedding will be helpful since there will
be both positive and negative reviews.
The DeepLSGR model is implemented using Tensorflow and keras layers for the
LSTM and GRU at the backend. The integration is carried out with the two stack lay-
ers of LSTM and GRU alternatively where each single cell has 16 neurons, learning
rate of 0.002 and number of epochs is 500.

13
 Downloaded from [Link]
[Link] [Link]

8720 Multimedia Tools and Applications (2023) 82:8709–8728

5.2 Baseline models

The baseline models that are considered for the performance evaluation of the proposed
model are listed as follows.

(i) Offset: It is an estimator used for the prediction of the ratings in recommendation
systems. It takes the average across all the ratings from the training set and predicts
the ratings in the test set.
(ii) Probabilistic matrix factorization (PMF) [10]: It is a method used for the prediction
of the ratings that works well for the sparse datasets. It is based on the probability
of the factors used for recommendations.
(iii) Collaborative Topic Regression (CTR) [21]: It gives the predictions using the topics
in the recommendations. It is also based on the probability of the topics aligned for
the prediction of the ratings.
(iv) LSTM and GRU: The proposed hybrid model is based on the LSTM [18] and GRU
[20]. The performance evaluation is carried out with only LSTM and GRU sepa-
rately.

5.3 Predicted ratings using DeepLSGR model

The experiments conducted with the DeepLSGR model using the two datasets and com-
pared with Offset, LSTM [18], GRU [20], PMF [10] and CTR [21] as the baseline works.
In addition, it was also compared with RNN having only LSTM and GRU separately. A
sample of the prediction of the ratings for a set of reviews is as shown in the Table 5. The
Table 5 gives an insight on a couple of new random test review texts taken from both the
datasets with their actual ratings obtained as a sample randomly and its predicted ratings
under each model. The proposed DeepLSGR model is capable of predicting ratings very
accurately as compared to the other methods. The least scorer is Offset because it takes
the average of ratings for the prediction. The predicted ratings of the LSTM and PMF are
closer because of the probability factors considered for the prediction of the ratings. Also,
the predicted ratings of the GRU and CTR are closer enough because of the topics consid-
ered for the prediction. However, the proposed DeepLSGR model outperforms all the other
models and gives better accurate predictions. The integration of the LSTM and GRU in the
hidden layers provided the improvement in ratings as the two stack layers of LSTM and
GRU alternatively provide improved connections between the sequences of the reviews.

5.4 Accuracy for DeepLSGR model

The performance evaluation of the DeepLSGR model was estimated using the RMSE and
accuracy using the same baseline works as discussed in the previous section. RMSE is
one of the parameters that is used to evaluate the loss of the DL model. A value of RMSE
close to 1 is considered as a better approximation model. The evaluation results of the com-
parison of the models are as shown in the Table 6. It can be observed from the results
in Table 3 that the RMSE values of the Offset, PMF, CTR, LSTM and GRU methods
are 1.172, 0.983, 0.976, 0.943, and 0.932 respectively for the Amazon Fine food reviews
dataset. The proposed DeepLSGR model has the RMSE value of 0.876 which is better as

13
 Table 5  Predicted ratings of DeepLSGR model
Review Actual rating Predicted ratings
Downloaded from [Link]

LSTM GRU​ Offset PMF CTR​ DeepLSGR

If Keurig brewers were evaluated by this particular K-Cup, Keurig would be out of business. Wolf- 1.0 0.9673 0.9563 0.9351 0.9632 0.9542 0.9847
gang Puck—French Roast Melange Francais is inexcusably terrible coffee
Multimedia Tools and Applications (2023) 82:8709–8728

This stuff sounded perfect…way too much red dye and no taste according to me and my 3 friends 2.0 1.9132 1.8367 1.842 1.9463 1.9637 1.9863
who tried it…kind of salty. Basically, I would stay away from it. I am tempted
I prefer the extra bold coffees, so this one really lacking on the flavor. It’s fairly weak, at the smallest 3.0 2.9635 2.9374 2.6352 2.9438 2.9727 2.9984
cup setting its drinkable, but I won’t be buying it again. I’d recommend
I really like the flavors and the fact that the chips are low in fat. I especially like the bbq and salt and 4.0 3.7634 3.8732 3.4274 3.9537 3.9728 3.9836
vinegar. I don’t care for the salt and pepper. I wish I could get a multipack
I loved this tea! It is refreshing and delicious with or without sweetener. A great buy! I would defi- 5.0 4.9682 4.9756 4.9142 4.9637 4.9758 4.9957
nitely recommend this product!
[Link]

13
8721
[Link]
 8722

13
Table 6  RMSE and accuracy of DeepLSGR model
Downloaded from [Link]

Models Dataset

Amazon Fine Food Reviews OpinRank

RMSE Training accuracy Validation accuracy Test accuracy RMSE Training accuracy Validation accuracy Test accuracy

Offset 1.172 70.85% 69.94% 68.95% 1.762 71.66% 70.59% 70.82%

PMF 0.983 71.78% 68.61% 69.38% 1.452 75.78% 74.68% 74.95%
CTR​ 0.976 72.83% 69.83% 67.46% 1.347 77.73% 75.85% 76.61%
LSTM 0.943 75.87% 71.82% 71.93% 1.218 79.87% 77.63% 78.73%
GRU​ 0.932 79.53% 76.67% 73.53% 1.207 81.66% 85.47% 86.73%
DeepLSGR 0.876 89.13% 95.42% 94.48% 0.982 91.27% 96.52% 97.56%
[Link]

Multimedia Tools and Applications (2023) 82:8709–8728

[Link]
 Downloaded from [Link]
[Link] [Link]

Multimedia Tools and Applications (2023) 82:8709–8728 8723

compared to the existing methods. Similarly, the RMSE values of the Offset, PMF, CTR,
LSTM and GRU methods are 1.762, 1.452, 1.347, 1.218 and 1.207 respectively for the
OpinRank dataset.
The offset method with the RMSE of 1.172 is the poorest estimator as it makes predic-
tions by not considering the differences among the users and items. The PMF model also
does not yield satisfactory results as the dataset is sparse and it requires complex modelling
of relationships between the users and items with the RMSE of 0.983. CTR also fails to
identify the potential words in the reviews as it is based on Bag of words approach and thus
the value of RMSE 0.976 which is high compared to the DeepLSGR model. The LSTM
and GRU alone also give better performance as compared to the other baseline models.
The proposed DeepLSGR model has the RMSE value of 0.982 which is better as com-
pared to the existing methods. The RMSE value of the proposed DeepLSGR model is bet-
ter as compared to the existing methods because of the bi-directional implementation of
the model using the LSTM and GRU as the hidden layers.
The comparison of the accuracy of the DeepLSGR model with the existing methods is
as shown in the Fig. 3. The testing accuracy of the DeepLSGR model is in proportion to
the validation accuracy for the Amazon fine food reviews dataset. Each method is run for
500 epochs as the number of epochs is one of the influential parameters for the comparison
of the accuracy. The convergence of the accuracy is measured as the epoch is reached 500.
The accuracy of the DeepLSGR model is increased by 3.08 times for the OpinRank dataset
as compared to the Amazon Fine food reviews. The accuracy of the model is still retained
in the case of OpinRank dataset which has sparse data in nature. It is clear Fig. 3 that the
PMF and CTR methods that is based on the Bag-of-words approach does not yield better
results as compared to the DeepLSGR model. The accuracy of the LSTM and GRU neural

Accuracy of DeepLSGR model

100.00% 96.52% 97.56%
95.42% 94.48%
89.13% 91.27%
90.00%
80.00%
70.00%
60.00%
Offset
Accuracy

50.00%
PMF [39]
40.00% CTR [40]
30.00% LSTM [44]
20.00% GRU [45]
10.00% DeepLSGR

0.00%
Training Validaon Test Training Validaon Test
Accuracy Accuracy Accuracy Accuracy Accuracy Accuracy
Amazon Fine Food Reviews OpinRank
Datasets

Fig. 3  Accuracy of DeepLSGR model

13
 Downloaded from [Link]
[Link] [Link]

8724 Multimedia Tools and Applications (2023) 82:8709–8728

networks are better compared to the offset, PMF and CTR models because of the sequence-
to-sequence encoding nature in the case of LSTM and GRU. However, the accuracy of the
DeepLSGR model is more compared to all the models because of the number of neurons
in the stacked layers of integration. The two stack layers of LSTM and GRU in the Deep-
LSGR are arranged alternatively where each single cell has 16 neurons and thereby provid-
ing the improved accuracy.

5.5 Performance comparison of DeepLSGR with data sparsity

The recommendation performance with data sparsity of the DeepLSGR model is meas-
ured using the recall metrics as shown in the Table 7. Recall@20 denotes the proportion of
cases having the desired item amongst the top-20 items in all test cases. It can be observed
from the results in Table 4 that the recall values of the Offset, PMF, CTR, LSTM and GRU
methods are 0.17, 0.32, 0.37, 0.43 and 0.46 respectively for the Amazon Fine food reviews
dataset which is 27% sparse. Similarly, the recall values of the Offset, PMF, CTR, LSTM
and GRU methods are 0.21, 0.34, 0.45, 0.47 and 0.53 respectively for the OpinRank data-
set which is 37% sparse.
The recall results of the DeepLSGR model for the Amazon Fine Food reviews data-
set are as shown in the Fig. 4 that provides better visualization and interpretability. The
results of the recall as shown in the Fig. 4 demonstrate that the offset method is the poorest
performer with recall percentage as 17–20% only. The PMF and the CTR methods have
recall of 32–34% and 33–37%. The LSTM and GRU methods show a recall of 37–43%
and 41–46% respectively. However, the DeepLSGR model has recall 57% which is higher
than the other models. The recall results of the DeepLSGR model for the OpinRank data-
set are as shown in the Fig. 5 that provides better visualization and interpretability. The
results of the recall as shown in the Fig. 5 demonstrate that the offset method is the poorest
performer with recall percentage as 15–21% only. The PMF and the CTR methods have
recall of 21–34% and 38–45%. The LSTM and GRU methods show a recall of 43–47% and
46–53% respectively. However, the DeepLSGR model has recall 61% which is higher than
the other models.
Recall is an important performance metric that need to be analyzed for the recom-
mendations. It gives the proportion of the recommendations in testing portions. The
proposed DeepLSGR model has the recall value of 0.57 and 0.61 for the Amazon Fine
food reviews dataset and OpinRank dataset respectively which is better as compared to
the existing methods. The increase in the recall is accounted for the ratings prediction

Table 7  Recall for DeepLSGR Model Amazon Fine Food OpinRank

model recommendations Reviews (37% Sparsity)
(27% Sparsity)
Recall @20 Recall @20

Offset 0.17 0.21

PMF 0.32 0.34
CTR​ 0.37 0.45
LSTM 0.43 0.47
GRU​ 0.46 0.53
DeepLSGR 0.57 0.61

13
 Downloaded from [Link]
[Link] [Link]

Multimedia Tools and Applications (2023) 82:8709–8728 8725

Fig. 4  Recall of DeepLSGR model for Amazon Fine Food Reviews

Fig. 5  Recall of DeepLSGR model for OpinRank

using the DeepLSGR model and then followed by the recommendations. The learning
rate of 0.002 and a dropout of 0.5 in the stacked layers improved the recall of Deep-
LSGR as compared to the other methods.

13
 Downloaded from [Link]
[Link] [Link]

8726 Multimedia Tools and Applications (2023) 82:8709–8728

6 Conclusion and future work

In RS both the ratings and the reviews play an important role to provide better and rank
recommendations. The reviews written by the user can reveal some of the information
related to the customer rating and behaviour aspects. In this paper, a deep learning model
DeepLSGR was proposed that uses the textual review information for recommender sys-
tems. It is a RNN with two hidden layers one for LSTM and another one for GRU. Dee-
pLSGR model extracts the reviews given by different users and converts it into ratings. It
is accomplished using a RNN based model with LSTM and GRU coupled with each other
as the hidden layers in RNN. The model is trained to estimate the ratings with minimum
prediction error and then CF is carried using neighbourhood classification. The experimen-
tal results with Amazon fine food dataset and OpinRank dataset reveal that the proposed
model has low RMSE (0.87%) and high recall values (61%) as compared to the existing
works. The results also indicate that the data sparsity problem can be solved using the pro-
posed DeepLSGR model as the recall rate is higher in the OpinRank dataset (37%) than the
Amazon fine food dataset (27%). In future, the work can be extended with more regulariza-
tion techniques within the hidden layers and improve the accuracy and recommendations.
It can also be used to design an explainable agent for recommendations.

References
1. Abdi MH, Okeyo G, Mwangi RW (2018) Matrix factorization techniques for context-aware collabora-
tive filtering recommender systems: a survey.
2. Alashkar T, Jiang S, Wang S, Fu Y (2017) Examples-rules guided deep neural network for makeup
recommendation. In: Thirty-first AAAI conference on artificial intelligence.
3. An HW, Moon N (2019) Design of recommendation system for tourist spot using sentiment
analysis based on CNN-LSTM. J Ambient Intell Humaniz Comput. [Link]
s12652-​019-​01521-w
4. Blei DM, Ng A, Jordan M (2003) Latent Dirichlet allocation. J Mach Learn Res 3:993–1022
5. Chen K, Franko K, Sang R (2021) Structured model pruning of convolutional networks on tensor pro-
cessing units. arXiv preprint. [Link]
6. Diao Q, Qiu M, Wu CY, Smola AJ, Jiang J, Wang C (2014) Jointly modeling aspects, ratings and senti-
ments for movie recommendation (JMARS). In: Proceedings of the 20th ACM SIGKDD international
conference on knowledge discovery and data mining. ACM, pp 193–202
7. Dong X, Yu L, Wu Z, Sun Y, Yuan L, Zhang F (2017) A hybrid collaborative filtering model with deep
structure for recommender systems. In: Thirty-first AAAI conference on artificial intelligence.
8. Ganesan KA, Zhai CX. “Opinion-based entity ranking”, information retrieval.
9. Gavilan D, Avello M, Martinez-Navarro G (2018) The influence of online ratings and reviews on hotel
booking consideration. Tour Manag 66:53–61
10. Gong X, Huang X (2019) A probabilistic matrix factorization recommendation method based on deep
learning. J Phys Conf Ser 1176(2):022043
11. Hwangbo H, Kim YS, Cha KJ (2018) Recommendation system development for fashion retail e-com-
merce. Electron Commer Res Appl 28:94–101
12. Jia X, Li X, Li K, Gopalakrishnan V, Xun G, Zhang A (2016) Collaborative restricted Boltzmann
machine for social event recommendation. In: 2016 IEEE/ACM international conference on advances
in social networks analysis and mining (ASONAM). IEEE, pp 402–405
13. Jiang L, Cheng Y, Yang L, Li J, Yan H, Wang X (2019) A trust-based collaborative filtering algorithm
for E-commerce recommendation system. J Ambient Intell Humaniz Comput 10(8):3023–3034
14. Kim KW, Park DH (2018) Emoticon by emotions: the development of an emoticon recommendation
system based on consumer emotions. J Intell Inf Syst 24(1):227–252
15. Kluver D, Ekstrand MD, Konstan JA (2018) Rating-based collaborative filtering: algorithms and eval-
uation. In: Social information access. Springer, Cham, pp 344–390

13
 Downloaded from [Link]
[Link] [Link]

Multimedia Tools and Applications (2023) 82:8709–8728 8727

16. Li S, Kawale J, Fu Y (2015) Deep collaborative filtering via marginalized denoising auto-encoder. In:
Proceedings of the 24th ACM international on conference on information and knowledge management.
ACM, pp 811–820
17. Liu Y, Zhang M (2018) Neural network methods for natural language processing.
18. Li Y, Liu T, Jiang J, Zhang L (2016) Hashtag recommendation with topical attention-based LSTM.
Coling
19. Li S, Zhao Z, Liu T, Hu R, Du X (2017) Initializing convolutional filters with semantic features for
text classification. In: Proceedings of the 2017 conference on empirical methods in natural language
processing. pp 1884–1889
20. Liu J, Wu C, Wang J (2018) Gated recurrent units based neural network for time heterogeneous feed-
back recommendation. Inf Sci 423:50–65
21. Ma C, Kang P, Wu B, Wang Q, Liu X (2019) Gated attentive-autoencoder for content-aware recom-
mendation. In: Proceedings of the twelfth ACM international conference on web search and data min-
ing. pp 519–527
22. McAuley J, Leskovec J (2013) From amateurs to connoisseurs: modeling the evolution of user exper-
tise through online reviews. WWW​
23. Mahata SK, Das D, Bandyopadhyay S (2019) Mtil 2017: machine translation using recurrent neural
network on statistical machine translation. J Intell Syst 28(3):447–453
24. McAuley J, Leskovec J (2013) Hidden factors and hidden topics: understanding rating dimensions with
review text. In: Proceedings of the 7th ACM conference on recommender systems. ACM, pp 165–172
25. Mukherjee S, Popat K, Weikum G (2017, June) Exploring latent semantic factors to find useful prod-
uct reviews. In: Proceedings of the 2017 SIAM international conference on data mining. Society for
Industrial and Applied Mathematics, pp 480–488
26. Nguyen VD, Sriboonchitta S, Huynh VN (2017) Using community preference for overcoming sparsity
and cold-start problems in collaborative filtering system offering soft ratings. Electron Commer Res
Appl 26:101–108
27. Okura S, Tagami Y, Ono S, Tajima A (2017) Embedding-based news recommendation for millions of
users. In: Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery
and data mining. ACM, pp 1933–1942
28. Parvina H, Moradi P, Esmaeilib S, Jalilic M (2018) An efficient recommender system by integrat-
ing non-negative matrix factorization with trust and distrust relationships. In: 2018 IEEE data science
workshop (DSW). IEEE, pp 135–139
29. Pennington J, Socher R, Manning CD (2014) GloVe: global vectors for word representation.
30. Pouyanfar S, Sadiq S, Yan Y, Tian H, Tao Y, Reyes MP, Shyu ML, Chen SC, Iyengar SS (2018) A sur-
vey on deep learning: algorithms, techniques, and applications. ACM Comput Surv (CSUR) 51(5):92
31. Seo S, Huang J, Yang H, Liu Y (2017) Interpretable convolutional neural networks with dual local and
global attention for review rating prediction. In: Proceedings of the eleventh acm conference on recom-
mender systems. ACM, pp 297–305
32. Shi Y, Larson M, Hanjalic A (2014) Collaborative filtering beyond the user-item matrix: a survey of
the state of the art and future challenges. ACM Comput Surv (CSUR) 47(1):3
33. Su X, Khoshgoftaar TM (2009) A survey of collaborative filtering techniques. Adv Artif Intell. [Link]
doi.​org/​10.​1155/​2009/​421425
34. Tan YK, Xu X, Liu Y (2016) Improved recurrent neural networks for session-based recommendations.
In: Proceedings of the 1st workshop on deep learning for recommender systems. ACM, pp 17–22
35. Thakkar P, Varma K, Ukani V, Mankad S, Tanwar S (2019) Combining user-based and item-based col-
laborative filtering using machine learning. In: Information and communication technology for intel-
ligent systems. Springer, Singapore, pp 173–180
36. Wu Y, Ester M (2015, February) Flame: a probabilistic model combining aspect based opinion mining
and collaborative filtering. In: Proceedings of the eighth ACM international conference on web search
and data mining. ACM, pp 199–208
37. Wu CY, Ahmed A, Beutel A, Smola AJ (2016) Joint training of ratings and reviews with recurrent
recommender networks.
38. Wu S, Ren W, Yu C, Chen G, Zhang D, Zhu J (2016) Personal recommendation using deep recurrent
neural networks in NetEase. In: 2016 IEEE 32nd international conference on data engineering (ICDE).
IEEE, pp 1218–1229
39. Xue W, Li T, Rishe N (2017) Aspect identification and ratings inference for hotel reviews. World Wide
Web 20(1):23–37
40. Zhang QS, Zhu SC (2018) Visual interpretability for deep learning: a survey. Front Inf Technol Elec-
tron Eng 19(1):27–39
41. Zhang Q, Yang LT, Chen Z, Li P (2018) A survey on deep learning for big data. Inf Fusion 42:146–157

13
 Downloaded from [Link]
[Link] [Link]

8728 Multimedia Tools and Applications (2023) 82:8709–8728

42. Zhang L, Wang S, Liu B (2018) Deep learning for sentiment analysis: a survey. Wiley Interdiscip Rev
Data Min Knowl Discov 8(4):e1253
43. Zhang S, Yao L, Sun A, Tay Y (2019) Deep learning based recommender system: a survey and new
perspectives. ACM Comput Surv (CSUR) 52(1):1–38
44. Zheng L, Noroozi V, Yu PS (2017) Joint deep modeling of users and items using reviews for recom-
mendation. In: Proceedings of the tenth ACM international conference on web search and data mining.
ACM, pp 425–434

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.

13