Page 59 - IJOCTA-15-1

BSO: Binary Sailfish Optimization for feature selection in sentiment analysis

results, the SVM-based and RF-based SA mod- Cai et. al. 46 proposed a recommendation model
els outperformed the others which are DT-based, based on DL (i.e., DeepCGSR) which uses users’
KNN-based, and LR-based models by consider- comment sets and items as a collection to do
ing F-score. For negative reviews, SVM-based cross-grained SA that consubstantiates textual
and RF-based SA algorithms were computed as critique sentiments and rating matrix. This
0.87 and 0.80 in terms of F-score, respectively. method was used on the e-commerce datasets
For neutral reviews, SVM-based and RF-based from Amazon. According to the results of
SA algorithms were computed as 0.77 and 0.64 the experiments, DeepCGSR model outperformed
in terms of F-score, respectively. For positive re- LFM 47 by 14.113%, SVD++ 48 by 13.786%,
views, SVM-based and RF-based SA algorithms TOPICMF 49 by 9.92%, DeepCoNN 50 by 5.122%
were calculated as 0.94 and 0.90 in terms of F- and NARRE 51 by 2.765%.
score, respectively.

2.2.3. SA based on nature inspired algorithms

2.2.2. SA based on DL There are not many studies where nature-inspired
languages are used as feature selectors in SA. For
this reason, since this study is thought to add
The Attention-Centred Bidirectional CNN-RNN
innovation and richness to the literature, similar
Deep Model (ABCDM) model, which removes
studies are presented in this section.
both past and future contexts, using 2 indepen-
dent bidirectional LSTM together with GRU lay- 52
ers was created to perform SA of product re- Sumathi et al. proposed to use Artificial Bee
views on Twitter in the study of Basiri et al. 43 Colony (ABC) as feature selector to classify opin-
The proposed ABCDM model was studied on five ion classification on Internet Movie Database
reviews and three Twitter datasets in order to (IMDb). Na¨ıve Bayes (NB), Fuzzy Unordered
recognize sentiment polarity. According to the Rule Induction Algorithm (FURIA) and Rip-
results, despite focusing only on document-level ple Down Rule Learner (RIDOR) are used as
sentence analysis, ABCDM, which aims to obtain classifiers with ABC. Experimental results show
both long reviews, accomplished its purpose by that ABC-RIDOR provides the best result with
using short tweet polarity classification. 93.75% accuracy.

In the study of Li et al., 44 DL-based SA algo- In the study Wahyudi and Kristiyanti, 53 SVM
rithms called lexicon integrated 2-channel Convo- Algorithm-Based Particle Swarm Optimization
lutional Neural Network–Long Short Term Mem- (PSO) was used for SA of smartphone product
ory family models were developed. To formu- review from the web site www.gsmarena.com. Ac-
late sample data entered, the sentiment padding cording to the experimental results, an increasing
methodology was used. This method was also was obtained in accuracy with SVM-based PSO
used to develop the sentiment information’s pro- as from 82.00% to 94.50%.
portion in every review and was formulated Sen-
timent Analysis’ premium lexicon components. Yuvaraj and Sabari 54 applied the Binary Shuf-
fled Frog algorithm (BSFA) as a feature selec-
In the study of Bilen and Horasan, 45 the Long- tor in sentiment analysis on the Twitter dataset
Short Term Memory Network and different ML al- provided by Stanford University. Radial Basis
gorithms were used on the Turkish Customer Re- Function (RBF), k-NN, NB and Logistic Model
views dataset which includes 8.500 customer re- Tree (LMT) were used as classifiers. The results
views collected from various electronic stores and showed that the BSFA-RBF model outperformed
Stanford Large Movie Reviews (IMDB) dataset others based on accuracy score.
which contains 50.000 movie reviews (25.000 posi-
tives and 25.000 negatives). Also, Zemberek NLP In the study of Naz et al., 55 for classification
Library for Turkish Language and Regular Ex- of sentiments using ensemble-based classifiers, to
pression models were used for the preprocessing solve the feature subset selection problem, a hy-
phase. The proposed model outperformed the brid strategy of minimum redundancy and max-
other ML models both on the Turkish dataset and imum relevance (mRMR) and Forest Optimiza-
IMDB dataset as 90.59% and 89.02% in terms of tion Algorithm (FOA)-based feature selection was
accuracy. proposed. This study employed to the Blitzer
53