Attentional Sentiment and Confidence Aware Neural Recommender

Model

Lamia Berkani

, Lina Ighilaza

and Fella Dib

Department of Artificial Intelligence & Data Sciences, Faculty of Informatics, USTHB University, Algiers, Algeria

Department of Computer Science, Faculty of Informatics, USTHB University, Algiers, Algeria

Keywords: Recommender System, Deep Learning, Hybrid Sentiment Analysis, Word Embedding, Conﬁdence Matrix.

Abstract: One of the major problems of recommendation systems is the rating data sparseness and information

overload. To address these issues, some studies are leveraging review information to construct an accurate

user/item latent factor. We propose in this article a neural hybrid recommender model based on attentional

hybrid sentiment analysis, using BERT word embedding and deep learning models. An attention mechanism

is used to capture the most relevant information. As reviews may contain misleading information (" fake

good reviews / fake bad reviews "), a conﬁdence matrix has been used to measure the relationship between

rating outliers and misleading reviews. Then, the sentiment analysis module with fake reviews detection is

used to update the user-item rating matrix. Finally, a hybrid recommendation is processed by combining the

generalized matrix factorization (GMF) and the multilayer perceptron (MLP). The results of experiments on

two datasets from the Amazon database show that our approach significantly outperforms state-of-the-art

baselines and related work.

1 INTRODUCTION

Over the past two decades, recommendation systems

have been widely employed in many domains. With

the information overload problem, finding suitable

products on online platforms is increasingly

becoming difﬁcult for users. Recommender systems

(RSs) are used, for instance, by Amazon to suggest

preferred products for customers, by Facebook and

LinkedIn to recommend people and webpages to

connect and follow, by YouTube to suggest related

videos. RSs learn the interests of users through their

historical behavior data and predict user preference.

Recently, deep learning (DL) techniques have been

applied in RSs to solve the cold start and data

sparsity problems. To further improve the

recommendation effectiveness of traditional

algorithms, He et al. (2017) proposed the neural

collaborative filtering (NCF) model, which is the

widely used model in RSs. They modelled the user-

item assessment matrix using a multilayer feedback

neural network (MLP) to learn the nonlinear

relationship between users and items.

On the other hand, with the considerable

development of social media, users often post

reviews to express their preferences and feelings

about items. These comments are considered reliable

indicators that can reflect the overall satisfaction of

users, expressing their preferable, non-preferable or

neutral opinions. These opinions are very useful for

understanding people’s references on items (Osman

et al., 2021). Several research works are applying

sentiment analysis (SA) in RSs (Wankhade et al.,

2022; Dang et al., 2021; Berkani and Boudjenah,

2023). However, many malicious users often share

misleading information about items in order to

manipulate or trap recommendation systems. As

reviews often contain fake good or fake bad reviews,

Li et al. (2021) leveraged the latent interaction factor

between users and items by exploiting the

interactivity of review information. To reduce the

impact of misleading comments on the model, they

used the confidence matrix to measure the

relationship between rating outliers and misleading

reviews. Birim et al. (2022) explored the most

effective feature combination in fake review

detection including the features of sentiment scores,

topic distributions, cluster distributions and bag of

words.

We focus in this article on these same issues and

we propose a novel recommender model using a

hybrid sentiment-based model and detecting

Berkani, L., Ighilaza, L. and Dib, F.

Attentional Sentiment and Conﬁdence Aware Neural Recommender Model.

DOI: 10.5220/0012193000003598

In Proceedings of the 15th Inter national Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2023) - Volume 1: KDIR, pages 323-330

ISBN: 978-989-758-671-2; ISSN: 2184-3228

323

fakereviews. We aim to enhance the user-item rating

matrix by predicting missing ratings and correcting

inconsistent values by matching the users’ ratings

and their associated comments on the items. Our

sentiment model is based on BERT word embedding

and several combinations of deep learning models

are considered: RNN-LSTM / BiLSTM and CNN-

LSTM / BiLSTM. For the detection of fake reviews,

we were inspired by the same approach of Li et al.

(2022) based on the confidence matrix. By

exploiting the enhanced user-item evaluation matrix,

our system generates predictions using a hybrid

recommendation algorithm using the generalized

matrix factorization (GMF) and the MLP models.

Extensive experiments conducted on two datasets

demonstrated the effectiveness of our model

compared to the state-of-the-art approaches and

baselines.

The rest of this article is organized as follows:

Section 2 provides a literature review about DL and

sentiment-based RSs. Section 3 presents our

sentiment-based approach for the recommendation

of items. The results of experiments are presented

and analysed in Section 4. Section 5 concludes this

work with some future perspectives.

2 LITERATURE REVIEW

With the latest successes in a variety of domains

including computer vision, machine translation,

speech recognition and natural language processing,

DL models have been exploited, in the last few

years, in RSs. Current models mainly use deep

neural networks to learn user preferences on items

for recommendations. Two categories of models can

be distinguished (Ni et al., 2021): (1)

recommendation models using a single neural

network structure (Zhang et al., 2017), such as

Multi-Layer Perceptron (MLP), Convolutional

Neural Network (CNN), Recurrent Neural Network

(RNN) and its variants Long Short Term Memory

(LSTM) and Gated Recurrent Unit (GRU); and (2)

recommendation models that combine the attention

mechanism and neural networks (Hu et al., 2018).

To overcome the rating data sparseness, users’

reviews are used for rating prediction. These reviews

or comments can express users’ overall satisfaction

on the items through their preferred, non-preferred

or neutral opinions. According to Osman et al.

(2021) these opinions will provide a better

understanding of user preferences on items.

Recently, the application of SA in RSs has been the

focus of extensive research. There are currently

three approaches to address the SA problem

(Bhavitha et al., 2017): (1) lexicon-based techniques,

including dictionary-based and corpus-based

methods (Salas-Zárate et al., 2017); (2) machine-

learning-based techniques (Zhang and Zheng, 2016),

including traditional techniques and DL techniques;

and (3) hybrid approaches, combining machine

learning and lexicon-based approaches (Pandey et

al., 2017). Bhattacharya et al. (2022) discussed and

analysed recent developments and related works

providing an overview of the different aspects of

SA. Among these works we reference the following

as examples: Diao et al. (2014) integrated two

parallel neural networks, and developed DeepCoNN

that jointly models users and items through reviews,

where the two CNNs are connected by a shared layer

facilitated by factorization machines. ConvMF

leverages the information in user-contributed

reviews and integrates CNN into the matrix

factorization (Kim et al., 2016). Chen et al. (2018)

developed a neural attentional regression model with

review-level explanations (NARRE) taking the

review usefulness into consideration.

Furthermore, some approaches combining two

models, such as SVM-enhanced CNN (Xue et al.,

2016), CNN with RNN (Rehman et al., 2019), have

yielded improved results. Dang et al. (2021) applied

a SA in RSs based on hybrid DL models and

collaborative ﬁltering. According to the authors, the

system architecture can integrate a variety of

techniques including the pre-processing strategy,

hybrid DL models for SA and methods for RSs.

After studying the proposed approaches, we

noticed the lack of a neural recommendation model

based on hybrid sentiment analysis exploiting both

attention mechanisms and fake reviews. Therefore,

we are motivated to propose a novel approach that

brings these features together, demonstrating the

contribution of each component in improving the

performance and the quality of recommendation.

The following section presents the proposed model.

3 OUR ASCAD-Rec MODEL

The proposed ASCAD-Rec model is summarized as

two main modules, as shown in Figure 1: (1) The

rating matrix enhancement module, which adjusts

the original rating matrix by combining it with our

sentiment model, which is based on fake reviews

detection; and (2) the recommendation module,

which aims to predict the user's rating on a given

item using the neural hybrid filtering (NHF) model.

KDIR 2023 - 15th International Conference on Knowledge Discovery and Information Retrieval

324

Figure 1: The proposed ASCAD-Rec model.

3.1 Rating Matrix Enhancement

3.1.1 Attentional Sentiment-Based Model

The sentiment module is based on users’ reviews on

items and ratings assigned on a 1-5 rating scale.

Figure 2 illustrates the different steps of this model.

First, we performed a set of pre-processing

operations on the data in order to convert it into an

interpretable and understandable form by our

system. We converted all words to lower case,

removed punctuation and stop words, and extracted

the stems of the words using the Snowball Stemmer

algorithm. Stemming is reducing a word to its base

word or stem where the words of similar kind lie

under a common stem.

Then, the second step is the embedding layer.

Processing textual data requires a conversion to

numbers before using any machine learning model.

Several methods allow this conversion, including the

traditional one-hot encoding on categorical data.

Currently, word embedding allows a better

representation of textual data. The embedding layer

allows converting each word into a fixed length

vector of defined size in order to better represent

words with reduced dimensions. I can understand

the context of a word and thus have similar

embeddings for similar words. The output vector

represents the input of the neural network’s hidden

layer. In this work, we used BERT, the Bidirectional

Encoder Representations from Transformers

proposed by (Devlin et al. 2018).

Sentiment-based

Hybrid DL

Fake reviews

detection

Attentional Sentiment and Conﬁdence Aware Neural Recommender Model

325

Figure 2: Hybrid sentiment analysis module based on attention mechanism.

BERT is used to transform text data to word

embedding. It was used to create the feature

vectors.Different combinations of DL models were

considered CNN → BiLSTM; and CNN → LSTM;

RNN → Bi-LSTM and RNN→ LSTM). These

models are followed by an attention layer to capture

only the most relevant information. The attention

mechanism is used to calculate the attention weight

for each component of the input sequence. The latter

are used to assign weights to the sequence so that the

most important components receive a higher weight

and therefore have a greater impact on the output.

The attention layer returns the context vector

representing the weighted sum of the sequence

elements, where the weights are the attention

weights. The attention mechanism takes as input the

output of the previous model. The result is used to

perform the final classification of the comments. It is

connected to a dense layer whose activation function

is "sigmoid". In order to train and validate the

sentiment analysis results, the reviews were labelled

with the sentiment score of the review, which can be

either positive or negative (value 1 or 0). Finally, the

output layer presents the result of the final rating of

an item i generated from the attentional sentiment

model.

3.1.2 Fake Reviews Detection

Misleading reviews are mainly caused by fake good

and fake bad reviews. When a user rates an item too

low (resp. too high) compared to his/her average low

ratings (resp. average high rating) compared to the

item's average rating, it will be considered as an

outlier value. Therefore, these ratings are given low

weights.

The confidence matrix is considered as a

regularization that aims to adjust the opinions,

taking into account the general opinion. It calculates

the difference between the score given by the user to

the item and the average of scores given by the user

to other items. Then it calculates the difference

between the score given by the user to the item and

the average of scores given to the item by other

users. If the difference is greater than a given

threshold (deviation rate), then it will be considered

as an outlier. The confidence level function used is

the same proposed in (Li et al., 2021).

3.1.3 Sentiment and Fake Reviews Detection

The sentiment score is combined with the fake

review detection module. The results obtained from

both the sentiment model and the fake reviews

detection model will be used to adjust the initial

rating matrix. Each initial rating will be combined

with the predicted rating using the following

weighted formula:

Score



α∗ Score







1α



∗Score



(1)

where:

 Score



: is the final score; Score



: is the

predicted score using SA with fake reviews

detection;

KDIR 2023 - 15th International Conference on Knowledge Discovery and Information Retrieval

326

 Score



: is the real score assigned by the

user on the item; and

 α: is a parameter used to adjust the

importance of each term of the equation.

3.2 Neural Hybrid Recommendation

The NHF model proposed in a previous work

(Berkani et al., 2019) was exploited for the

recommendation. NHF is an extension of the very

popular NCF model (He et al., 2017), which

concatenates the results of the two models GMF and

MLP, then passes the result through a dense layer

with a "Sigmoid" activation function to obtain the

prediction result.

The GMF model considers the latent user and

item factors as input, representing the hidden

characteristics that determine user preferences and

item features. These vectors are given as input to a

neural layer which performs the product of the two

vectors. The result is then sent to the predictive layer

represented by a dense layer with a "sigmoid"

activation function. On the other hand, the vectors of

the user and item latent factors, concatenated with

their respective information, are passed through an

MLP network, providing as output the prediction of

the evaluation score.

3.3 Training

The data extracted from the datasets must be pre-

processed and transformed before passing through

the different models. Then, these data are divided

into two sets considering 80% of the data for

training and the rest for testing in order to evaluate

the performance of our models. To guarantee a good

performance of our models, we used the "Binary

Cross-Entropy" cost function. In order to improve

the performance of our models, we use the Adam

(Adaptive Moment estimate) algorithm, a stochastic

optimization method (Kingma and Ba, 2014). This

algorithm will minimize the cost function, as a lower

loss means a better performance of our models.

4 EXPERIMENTS

We present in this section the experiments carried

out to evaluate the hybrid SA including the

contribution of the attention and fake review model.

Then we evaluated the hybrid recommender model

by comparing its performance with some exiting

models. Finally, we compared our approach with the

state-of-the art recommender models.

4.1 Datasets and Evaluation Metrics

For the training and evaluation of our sentiment

analysis models, we used the IMDB database, which

includes 50,000 reviews classified into two

categories (positive and negative comments). For

our experiments, we used a dataset from the Yelp

database, considering the shopping category. This

dataset, with a density of 0.25%, comprises 2,472

users, who have carried out 55,738 reviews on 8,785

items.

For the experiments, we used the Mean Absolute

Error (MAE) and the Root Mean Square Error

(RMSE) evaluation metrics. MAE and RMSE have

been used as they are the most popular predictive

metrics to measure the closeness of predictions

relative to real scores.

4.2 Evaluation Results

4.2.1 Evaluation of Sentiment Models

To generate the embedding vectors for each

comment, we performed the pre-processing of the

texts using the Python libraries (tokenization,

removal of empty words and punctuation, encoding,

padding, etc.). Then we generated the embedding

vectors using the BERT model. Next, we varied the

hyper-parameter values.

We varied the number of neurons in the last

dense layer of our models. We can notice from

Table 1 that 128 performed better than the other

models, with the exception of the hybrid LSTM-

RNN model, which requires 32 neurons.

Table 1: Evaluation of hybrid SA models varying the number of neurons.

LSTM BiLSTM

CNN RNN CNN RNN

RMSE MAE RMSE

MAE

RMSE MAE RMSE MAE RMSE MAE RMSE MAE

0.2755 0.1375 0.3230 0.1336 0.2640 0.1059 0.2679 0.1135 0.2659 0.1140 0.2777 0.1110

0.2811 0.1354 0.2824 0.1282 0.2777 0.1178 0.2897 0.1217 0.2740 0.1099 0.2707 0.1149

128

0.2770 0.1331 0.2746 0.1324 0.2653 0.1029 0.2708 0.1124 0.2607

0.1074 0.2710 0.1126

256

0.2855 0.1448 0.2797 0.1335 0.2649 0.1074 0.2724 0.1174 0.2651 0.1143 0.2835 0.1180

Attentional Sentiment and Conﬁdence Aware Neural Recommender Model

327

Table 2: Contribution of the attention mechanism on the performance of our models.

Models

Without attention With attention

MAE RMSE MAE RMSE

BiLSTM 0.1307 0.2817 0.1324 0.2746

BiLSTM-CNN 0.1202 0.2705 0.1074 0.2607

Table 3: Variations of PF and MLP Layers of the hybrid model.

4 5 6 7

RMSE MAE RMSE MAE RMSE MAE RMSE MAE

8 0.2037 0.1540 0.2042 0.1533 0.2052 0.1523 0.2064 0.1521

12 0.2056 0.1537 0.2051 0.1530 0.2067 0.1538 0.2052 0.1506

16 0.2061 0.1535 0.2031 0.1510 0.2062 0.1516 0.2052 0.1513

18 0.2050 0.1532 0.2051 0.1522 0.2064 0.1522 0.2065 0.1509

Table 4: Variations of α.

Models

NHF GMF PMF SVD++

RMSE MAE RMSE MAE RMSE MAE RMSE MAE

Without SA 0.2031 0.1510 0.2318 0.1689 2.1399 1.7754 0.9997 0.7684

With

α = 0.3 0.2012 0.1491 0.2280 0.1785 2.1136 1.7482 0.9818 0.7632

α = 0.4 0.2074 0.1547 0.2335 0.1848 2.0574 1.7207 1.0085 0.7916

α = 0.6 0.2327 0.1743 0.2537 0.2053 2.0982 1.7764 1.1116 0.8887

α = 0.7 0.2501 0.1879 0.2676 0.2182 2.1387 1.8052 1.1929 0.9580

We can observe that the BiLSTM-CNN model

performed better than the other models. The

BiLSTM network takes into account the context

before and after each word in the sequence,

providing a better understanding of the overall

context. Convolution layers are particularly useful

for capturing local features.

We have demonstrated the contribution of the

attention mechanism on the performance of the

BiLSTM and BiLSTM-CNN models. The results

presented in Table 2 show that the attention

mechanism considerably improves the predictions of

our models.

4.2.2 Recommendation Evaluation

We varied the vector size (T.E) of the hybrid model

(NHF), considering T.E equal to 8, 16, 32 and 64.

The best results were obtained with a T.E equal to 8

and a performance in terms of MAE equal to 0.1510

and RMSE equal to 0.2031.

Furthermore, as presented in Table 3, we varied the

number of MLP layers and the predictive factor (PF)

of the MLP network of the NHF model. Given that

the PF represents the number of neurons in the last

layer.

We can notice that the best results were obtained

with PF equal to 16 and a number of layers equal

to5. This configuration will be considered in the rest

of our experiments. Then, we varied the parameter

α, which adjusts the balance between the original

and the predicted ratings, of the NHF hybrid model.

We compared its performance with three simple

matrix factorization models GMF, PMF and SVD++

(see Table 4). We can see from the results that the

NHF, GMF and SVD++ models perform better with

α=0.3, while the PMF model performed better with

α = 0.4. Moreover, the results proved the

contribution of SA on NHF compared to the model

without SA.

To assess the contribution of fake reviews

detection on our model, we have shown in the table

a performance comparison between the model with

fake reviews (ASCAD-Rec) and the model without

fake reviews (ASAD-Rec). We obtained the

performance of MAE = 0.1438 and RMSE = 0.1964

with the ASCAD-Rec model, while with the ASAD-

Rec model, we obtained the performance of MAE =

0.1491 and RMSE = 0.2012. These results

demonstrate that the detection of fake reviews and

the regularization of the sentiment score improve the

performance of our model and enable more efficient

learning.

4.2.3 Comparison with Related Work

We present in this sub-section a comparison of our

approach with existing work (see Table 5):

• PMF: is a machine learning model based on

probabilistic principles for matrix factorization.

• SVD++: is a matrix factorization technique

based on singular value decomposition.

KDIR 2023 - 15th International Conference on Knowledge Discovery and Information Retrieval

328

• NCF: is a matrix factorization method based on

deep learning techniques for collaborative

filtering.

• DeepCoNN: is a deep learning-based approach

that exploits user comments to generate

recommendations.

• NARRE: exploits user comments by integrating

an attention mechanism to create vectors of

latent user and item factors.

Table 5: Performance comparison of different models.

Models MAE RMSE

PMF 1.7207 2.0574

SVD++ 0.7632 0.9818

NCF 0.1689 0.2318

NAREE 0.8767 1.1360

Dee

CoNN 0.6370 0.8192

NHF 0.1510 0.2031

ASAD-REC 0.1491 0.2012

ASCAD-Rec 0.1438 0.1964

The results obtained show the effectiveness of

our approach, compared with other state-of-the-art

models. Furthermore, the sentiment-based

recommender model with fake reviews detection

outperformed the model without fake reviews

detection.

4.3 Overall Discussion of Our Results

The experiments carried out on our approach using

the shopping yelp dataset demonstrated the impact

of sentiment analysis and fake review detection on

the effectiveness of the recommendation. We

obtained an MAE equal to 0.1438, compared with an

MAE equal to 0.1491 for the model without fake

reviews detection.

Experiments carried out on the sentiment

analysis model demonstrated the importance of

combining DL models with the significant

contribution of the attention mechanism in

improving the sentiment predictions. The model

with the attention mechanism yielded a MAE equal

to 0. 1074, compared with a MAE equal to 0.1202

obtained with the model without the attention

mechanism. Furthermore, it should be noted that the

hyper-parameter values considerably affect the

performance of the models.

5 CONCLUSION

In this article, we have presented a neural hybrid

recommendation model that combines the

Generalized Matrix Factorization (GMF) and the

Multilayer Perceptron (MLP) models to learn

respectively the linear and the nonlinear relationship

between users and items. This model leverages a

user-item rating matrix that has been adjusted by a

sentiment model to predict the sentiment scores

through users' textual comments on items. To further

improve our sentiment model, we used the attention

mechanism to capture the most important

information, and we identified fake reviews using a

confidence matrix. The results of our experiments

carried out on the Yelp shopping dataset

demonstrated the effectiveness of our model, which

outperformed the state-of-the-art models. Our

evaluations demonstrated the contribution of the

hybrid sentiment model compared to the simple

models, and the added value of the attention

mechanism and the detection of fake reviews. On the

other hand, our hybrid model performs better than

several factorization matrix-based models.

In our future work we plan to carry out further

experiments on larger datasets with higher density,

use multilingual comments and explore other deep

learning models such as the transformers.

REFERENCES

Al-Saqqa, S., Awajan, A. (2019). The use of word2vec

model in sentiment analysis: A survey. In Proceedings

of the 2019 International Conference on Artificial

Intelligence, Robotics and Control, pp. 39-43.

Berkani, L. Kerboua, I. Zeghoud, S. (2020).

Recommandation hybride basée sur l’apprentissage

profond. In Proceedings of EDA Conference, Revue

des Nouvelles Technologies de l’Information, RNTI

B.16, pp. 69–76. ISBN: 979-10-96289-13-4.

Berkani, L., Zeghoud, S., Kerboua. I. 2022. Chapter 19 -

Neural hybrid recommendation based on GMF and

hybrid MLP, Editor(s): Rajiv Pandey, Sunil Kumar

Khatri, Neeraj kumar Singh, Parul Verma, Artificial

Intelligence and Machine Learning for EDGE

Computing, Academic Press, 287-303,

https://doi.org/10.1016/B978-0-12-824054-0.00030-7.

Berkani, L., Boudjenah, L. (2023). S-SNHF: sentiment

based social neural hybrid filtering. Int. J. Gen.

Syst. 52(3), pp. 297-325

Bhattacharya, S., D., Sarkar, D-K., Kole, Jana, P. (2022).

Chapter 9 - Recent trends in recommendation systems

and sentiment analysis”, Editor(s): Sourav De, Sandip

Dey, Siddhartha Bhattacharyya, Surbhi Bhatia, In

Hybrid Computational Intelligence for Pattern

Analysis, Advanced Data Mining Tools and Methods

for Social Computing, Academic Press, pp. 163-175.

Bhavitha, B.; Rodrigues, A.P.; Chiplunkar, N.N. (2017).

Comparative Study of Machine Learning Techniques

in Sentimental Analysis. In Proceedings of the 2017

Attentional Sentiment and Conﬁdence Aware Neural Recommender Model

329

International Conference on Inventive Communication

and Computational Technologies (ICICCT),

Coimbatore, India, 10–11 March 2017; IEEE: New

York, NY, USA, pp. 216–221.

Birim, S.O, Kazancoglu, I., Mangla, S. K., Kahraman, A.,

Kumar, S., Kazancoglu, Y. (2022). Detecting fake

reviews through topic modeling. Journal of Business

Research 149, 884–900

Chen, C., Zhang, M., Liu, Y., Ma, S. (2018). Neural

attentional rating regression with review-level

explanations. In Proceedings of the 2018 World Wide

Web Conference, pp. 1583-1592.

Dang, C.N., Moreno-García, M.N., De la Prieta, F. (2021).

An Approach to Integrating Sentiment Analysis into

Recommender Systems. Sensors 21, 5666,

https://doi.org/10.3390/s21165666.

Devlin, J., Chang, M-W., M-W., Lee, K., Toutanovan, K.

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

transformers for language understanding. ArXiv,

arXiv:preprint/04805.

Diao, Q., Qiu, M., Wu, C.Y., Smola, A.J., Jiang, J., Wang,

C. (2014). Jointly modeling aspects, ratings and

sentiments for movie recommendation. In Proceedings

of the 20th ACM SIGKDD International Conference

on Knowledge Discovery and Data Mining, pp. 193-

202.

Duantengchuan L., Liu, H., Zhang, Z., Lin, K., Fang, S.,

Li, Z., N-N. Xiong, N-N. (2021). CARM: Conﬁdence-

aware recommender model via review representation

learning and historical rating behaviourin the online

platforms. Neurocomputing 455, pp. 283-296.

He, X., Liao, L., Zhang, H., Nie, L., Hu, X., Chua, T-S.

(2017). Neural collaborative ﬁltering. In Proceedings

of the 26th Internet Conference on World Wide Web,

pp. 173–182.

Hu, B., Shi, C., Zhao, W.X., Yu, P.S. (2018). Leveraging

meta-path based context for top-n recommendation

with a neural co-attention model. In Proceedings of

the International Conference ACM SIGKDD, pp. 153-

1540.

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

Convolutional matrix factorization for document

context-aware recommendation. In Proceedings of the

10th ACM Conference on Recommender Systems

(RecSys), pp. 233-240.

Kingma DP, Ba J. (2014). Adam: A method for stochastic

optimization. In arXiv preprint arXiv: 1412.6980.

Li, D., Liu, H., Zhang, Z., Lin, K., Fang, S., Li, Z., Xiong,

N. (2021). CARM: Confidence-aware recommender

model via review representation learning and historical

rating behavior in the online platforms.

Neurocomputing 455, pp. 283–296

Mikolov, T., Chen, K., Corrado, G., Dean, J. (2013a)

Efficient Estimation of Word Representations in

Vector Space. ICLR - Workshop Poster,

Mikolov, T., Sutskever,I., Chen, K., Corrado, G.S., Dean,

J. (2013b). Distributed Representations of Words and

Phrases and their Compositionality. In Proceedings of

the 26th International Conference on Neural

Information Processing Systems NIPS'13, Volume 2,

Advances in Neural Info. Processing Systems 26.

Mnih, A., Salakhutdinov, R. (2007). Probabilistic matrix

factorization. In advances in neural information

processing systems.

Ni, J., Huang, Z., Cheng, J., Gao, S. (2021). An effective

recommendation model based on deep representation

learning. Information Sciences, Vol. 542, N °1, pp.

324-342.

Osman, N.A., Mohd Noah, S.A. Darwich, M., Mohd, M.

(2021). Integrating contextual sentiment analysis in

collaborative recommender systems. PLoS ONE Vol.

16, N° 3: e0248695,.

Pandey, A.C.; Rajpoot, D.S.; Saraswat, M. (2017). Twitter

sentiment analysis using hybrid cuckoo search

method. Info. Process. and Manag., 53, pp. 764–779.

Pennington, J., Socher, R., Manning, C-D. (2014). GloVe:

Global Vectors for Word Representation. Empirical

Methods. In Natural Language Processing (EMNLP),

https://nlp.stanford.edu/pubs/glove.pdf.

Rehman, A.U., Malik, A.K., Raza, B., Ali, W. (2019). A

hybrid CNN-LSTM model for improving accuracy of

movie reviews sentiment analysis. Multimedia Tools

and Applications 78, pp. 26597–26613.

Salas-Zárate, M.D.P., Medina-Moreira, J., Lagos-Ortiz,

K., Luna-Aveiga, H., Rodriguez-Garcia, M.A.,

Valencia-García, R. (2017). Sentiment analysis on

tweets about diabetes: An aspect-level approach.

Computational and Mathematical methods. pp. 1–9.

Wankhade, M., Sekhara Rao, A-C., Kulkarni, C. (2022). A

survey on sentiment analysis methods, applications,

and challenges. Artificial Intelligence Review.

Xue, D-X., Zhang, R., Feng, H., Wang, Y.-L. (2016).

CNN-SVM for microvascular morphological type

recognition with data augmentation. Journal of

Medical and Biological Engineering. 36, pp. 755–764.

Zhang, S., Yao, L., Sun, A. (2017). Deep learning based

recommender system: A survey and new perspectives.

In arXiv preprint arXiv: 1707.07435.

Zhang, X., Zheng, X. (2016). Comparison of Text

Sentiment Analysis Based on Machine Learning. In

Proceedings of the 2016 15th International

Symposium on Parallel and Distributed Computing

(ISPDC), Fuzhou, China, 8–10 July 2016; IEEE: New

York, NY, USA, pp. 230–233.

Zheng, L., Noroozi, V., Yu, P-S. (2017). Joint deep

modeling of users and items using reviews for

recommendation. In Proceedings of the tenth ACM

international conference on websearch and data

mining, pp. 425-434.

KDIR 2023 - 15th International Conference on Knowledge Discovery and Information Retrieval

330