Page 58 - IJOCTA-15-1

O. Ayana, D. F. Kanbak, M. Kaya Keles / IJOCTA, Vol.15, No.1, pp.50-70 (2025)

2.1. SA based on user-item ratings The results indicated that for the sentiment clas-
sification as positive, negative and neutral classes,
the accuracy rate was over 90%.
The goal of the classic recommendation algo-
rithms, which rely solely on user scores, is to dis- 37
cover user choices and propose information. 32,33 Rummeli et al. developed SA model using
Based on user-item ratings, both memory-based lexicon-based methods and ML algorithms for
and model-based techniques are frequently imple- Turkish texts. The dataset, which includes
mented to common filtering. There are 2 filter- 272.218 examples including the user review, score
ing techniques adopted in the memory-based ap- and product’s web page address, was created us-
38
proach: item filtering and item-item filtering. In ing the Hepsiburada e-commerce website. The
39
40
the first method a specific user is retrieved, and SentiTurkNet dictionary and Zemberek tool
similar users are found based on their ratings, were used in the data preprocessing stages. The
whereas in the latter one users that like a certain Cross-Verification (CV) method was used to eval-
item are identified and based on this other items uate 4 different ML algorithms including KNN,
that similar users like are found. An item-based NB, RF, and SVM algorithms by taking into ac-
collaborative filtering algorithm was created for count the accuracy, AUC value, F-score, and to-
customers’ preferences on Amazon, which pro- tal running time. According to the results, all
duces enormous revenue for Amazon by Linden models’ average accuracy were around 0.73 and
et al. 34 all models are able to obtain emotion descriptions
of new Turkish texts.

In the study of Sharma and Dutta, 35 a novel ap-
proach named as ”SentiDraw” was proposed. In In the study of Zhao et al., 41 the proposed Lo-
this new approach, Sentiment Orientation scores cal Search Improvised Bat Algorithm based El-
were calculated using star ratings and the proba- man Neural Network (LSIBA-ENN) algorithm to
bility distribution of terms in comments with sev- find polarity of online product criticisms was ap-
eral star ratings. The results showed that lexicons plied to App and Movie & TV datasets collected
built with SentiDraw technique has a great per- from Amazon. The proposed LSIBA-ENN’s per-
formance. According to the accuracy score, the formance was compared with SVM, NB, and ENN
success rate ranges from 78.0% to 81.6%. Further- algorithms using performance evaluation metrics
more, in the study of Sharma and Dutta, 35 a hy- which are precision, recall, F-score, and accuracy.
brid approach, in which SentiDraw was used with Focusing on TW methods such as Word 2vector,
supervised methods to provide the most advanced Term Frequency, Term Frequency-Inverse Docu-
performance, was also recommended to determine ment Frequency, Term Frequency-Distinguishing
reviews’ polarity. Feature Selector as well as Log Term Frequency-
based Modified Inverse Class Frequency abbre-
viated as LTFMICF, these methods’ mentioned
2.2. SA based on user reviews metrics’ results were examined. According to
the results, LSIBA-ENN achieved highest per-
formance level for both datasets. When using
In this section, studies on SA will be examined un-
LTFMICF, LSIBA-ENN gave 92.01% value as ac-
der three headings: SA based on ML, SA based on
curacy on the App dataset, while the accuracy
DL, and SA based on nature inspired algorithms.
of LSIBA-ENN was 93.91% on the Movie and
TV dataset. Thus, the study has confirmed that
LSIBA-ENN is effective ML to execute SA of
2.2.1. SA based on ML
product data with precise results.

In the study of Murugan and Devi, 36 data stream-
ing of Twitter based on SA was classified us- Demircan et al. 42 proposed new models using ML
ing hybridization methods. Three algorithms in- methods which are DT, KNN, Logistic Regres-
cluding the Decision Tree (DT), Particle Swarm sion (LR), RF, and SVM to predict emotions ex-
Optimization (PSO), and the Genetic Algorithm pressed through Turkish texts in social media.
(GA) were used for SA’s classification. Therefore, More than 250.000 reviews of different products
six hundred twitter data were collected with the from the hepsiburada.com website, which are di-
aid of feature generation and URL-based security vided into 3 classes as positive, negative and neu-
tool. A hybrid of PSO, GA and DT algorithms tral, and review scores are turned into a table to
was used to classify the sentiments for analysis. be used in ML-based SA models. According to the
52