\documentclass[research]{elsarticle} \usepackage{lineno,hyperref} \usepackage{graphicx} \usepackage{fixltx2e} \usepackage{mathtools} \usepackage{amsmath} \modulolinenumbers[5] \journal{Journal of \LaTeX\ Templates} \begin{document} \begin{frontmatter} \title{Opinion mining framework using proposed RB-Bayes model for text classification \tnoteref{abc}} %% Group authors per affiliation:[] \author[mymainaddress]{Rajni Bhalla\corref{mycorrespondingauthor}} \ead{rajni.b27@gmail.com} \author[mysecondaryaddress]{Dr.Amandeep } \ead{amandeep1@lpu.co.in} \cortext[mycorrespondingauthor]{Corresponding author} \address[mymainaddress]{Research Scholar, School of Computer Application, Lovely Professional University, India} \address[mysecondaryaddress]{Associate Professor , Lovely Professional University, India} \begin{abstract} Information mining is a capable idea with incredible potential to anticipate future patterns and conduct. It alludes to the extraction of concealed information from vast data sets by utilizing procedures like factual examination, machine learning, grouping, neural systems and genetic algorithms. In naïve baye’s, there exists a problem of zero likelihood. This paper proposed RB-Bayes method based on baye’s theorem for prediction to remove problem of zero likelihood. We also compare our method with few existing methods i.e. naive baye’s and SVM. We demonstrate that this technique is better than some current techniques and specifically can analyze data sets in better way. At the point when the proposed approach is tried on genuine data-sets, the outcomes got improved accuracy in most cases. RB-Bayes calculation having precision 83.333. %formatted \LaTeX\ manuscript. \end{abstract} \begin{keyword} \texttt{Naive Bayes}\sep RB-Bayes\sep SVM \sep hotencoder \MSC[2018] \end{keyword} \end{frontmatter} \section{Introduction} Information Discovery from Data (KDD) is the objective of information mining process \cite{1}.Everybody is active on social media and their motive is not only too active on social media but also to generate information. Before purchasing anything, we always check reviews on social media. That reviews not only useful for consumer but also for manufacturer. With this pattern, there are more examinations on programmed investigation and blend of data from client audits gathered from online networking. Because of the valuable data gave by these investigations, makers can enhance their items, the specialists can change strategies in like manner, and in addition clients can pick the item most appropriate to their conditions. They can improve the features and can increase sales of their product \paragraph The advancement of innovation alongside the request of investigating obstinate data has prompted another exploration subject in normal dialect preparing and information mining named "conclusion mining and notion examination". Concentrates on this issue began from the 2000s, tending to a few issues including extremity characterization \cite{2}-\cite{3}, subjectivity characterization \cite{4}-{7}, and conclusion spam location\cite{8}-\cite{9}\cite{10}. Early examinations concentrated on the basic data sources which for the most part contain the sentiment on one subject and the errand is the manner by which to arrange this supposition into the classes negative, unbiased or positive \cite{11}-\cite{12}-\cite{13}. Late issues with more entangled data sources have pulled in numerous specialists. An audit frequently contains assessments on various item perspectives, or contains practically identical feelings. A few issues of concern incorporate identifying equivalent sentences \cite{14}-\cite{15}, deciding viewpoints \cite{16}-\cite{17}-\cite{18}, rating angles \cite{19}-\cite{20}-\cite{21} or deciding viewpoint weights \cite{22}-\cite{23}-\cite{24}. Viewpoint based feeling examination as of late turns into a vital issue, in which we have to give the integrated slant on each item viewpoints. Viewpoint based feeling is more important because both manufacturer and customers want to know that which features are more popular and on which features they need to do improvement. For example: Before purchasing mobile phone or laptop, customer always ask for their feature like camera,video,Bluetooth,brand,product price, special offer, dual sim,charging hour, battery backup, operating system etc. After finding which features are important, manufacturer can improve on that particular aspect. \paragraph Some past examinations, for example, \cite{25}-\cite{26} have proposed a model called the Latent Rating Regression (LRR) which is a sort of Latent Dirichlet Allocation to break down both perspective appraisals and viewpoint weights, or \cite{26} utilized the Maximum A Posterior (MAP) procedure to handle the angle sparsity issue. This paper \cite{25} depicted a non-parametric Bayes technique to perform parallel order on graphs. In perspective of computational many-sided quality it may be more worthwhile to think about different strategies to adaptively discover the tuning parameter, for example, observational.\cite{26} misfortune is developed. An experimental probability deduction on the evaluated parameter vector is made. In this paper, our essential objective is to proposed new strategy to expel issue of probability of zero in naive bayes and to explore the estimation for parameters also, non-parametric capacity both theoretically and practically. The rest of paper is composed as takes after. In Section 2, we present the dataset. In Section 3, we present proposed RB-Bayes calculation. In Section 4, we speak to usage of RB-bayes calculation and furthermore demonstrated the examination with other algorithms. We give some finishing up comments and future work in Section 5. %\begin{itemize} %\item document style %\item baselineskip %\item front matter %\item keywords and MSC codes %\item theorems, definitions and proofs %\item lables of enumerations %\item citation style and labeling. %\end{itemize} \section{Related Work} This \cite{27} investigation proposed applicable element extraction calculation for X-beam restorative pictures and decided machine learning strategies for programmed X-beam medicinal picture characterization. This examination likewise assessed distinctive picture highlights (predominantly worldwide, neighbourhood, and joined) and classifiers.\cite{28} The best test result for music feeling grouping was the utilization of Random Forest strategies for verses and sound features.Some mixture method can be manufacture utilizing irregular backwoods moreover. \section{Proposed Methodology} \subsection{Dataset} From dataset, we can predict whether customer will purchase computer or not. We have five parameters. On basis of these parameters we will predict. Parameters are age, income, student and credit rating.As we can see from Figure 1 Age wise maximum response are from youth and senior. To check accuracy and for comparison we test proposed algorithm on small dataset. Similarly we have parameter income that consists of three values high, medium and low. Student type will be binomial consist of two values either yes or no.Credit rating will also be binomial type and consist of two values either fair or excellent. We want to predict variable buys computer consist of two values either yes or no. ; Similarly we have other parameters i.e. income' which consist of three values High, medium and low ; student which consist values – Yes or No ; Credit rating which consist values fair or excellent. \begin{figure} \centering \includegraphics[width=0.5\textwidth]{agewise.jpg} \caption{Agewise Response.} \end{figure} \subsection{RB-Bayes Algorithm} RB-Bayes is one of simplest supervised technique. It is a classification system in light of bayes theorem. It is mostly used in text classification. Naive bayes is also based on bayes theorem. But unable to handle problem of likelihood of zero possibility. RB-Bayes is proposed to solve this problem. RB-Bayes algorithm provides a way of calculating prediction. Look at equation below. \begin{equation} P_yF = \frac{T_y}{Total_Sampleset} * \frac{Ty_i+..........+Ty_n}{T_F * T_y} \end{equation} \subsubsection{RB-Bayes algorithm steps} \begin{enumerate} \item Each tuple that we wish to classify is represented by \begin{equation} X = (x_1,x_2......x_n) \end{equation} \item There are n numbers of labels. Given a tuple, X, the classifier will anticipate that X has a place with the label having the highest value from all labels. \item Checking highest value for labels \begin{equation} P_yF > P_nF \hspace{1cm} where y \ne n \end{equation}Value of y and n are different labels. Maximum value from all labels will do prediction \item Maximize P\textsubscript yF \begin{equation} Mean = \frac{T_y}{Total_SampleSet} \end{equation} \begin{equation} P_yF = Mean*\frac{T_ya+T_yb+T_yc+T_yd+......T_yn}{T_F * T_y} \end{equation} \item T(Y\textsubscript i), for i=1, 2, 3…......n, is a prior probability value depends on labels. Prior probability of each class can be computed based on training tuples. We calculate T\textsubscript ya ,T\textsubscript yb ,T\textsubscript yc and T\textsubscript yd………...........T\textsubscript yn and this needs to be maximized. \item T\textsubscript ya is calculated by comparing value with P(y).Count will store in T\textsubscript ya+T\textsubscript yb+T\textsubscript yc+T\textsubscript yd………….........+T\textsubscript yn wherever both values are active. Similarly for T\textsubscript yb ,T\textsubscript yc and T\textsubscript yd . \begin{equation*} \begin{aligned} T\textsubscript ya = \prod \limits_{k=1}^N P(x_k | Y_i) \\ =P(x_1 | Y_i) * P(x_2 | Y_i) * .............P(x_n | Y_i) \end{aligned} \end{equation*} \item Predicting class label depending upon highest value after comparing value of P\textsubscript yF ,P\textsubscript nF \item Factors affecting y and n can be n number of factors i.e.T\textsubscript ya ,T\textsubscript yb ,T\textsubscript yc and T\textsubscript yd………........T\textsubscript yn \item The classifier predicts that the class label of tuple X is the class P\textsubscript yF or P\textsubscript nF If and only if \begin{equation} \frac{T_y}{Total_ {ss}} * \frac{T_ya+T_yb+T_yc+T_yd+..........T_yn}{T_F * T_y} > or < \frac{T_n}{Total_{ss}}* \frac{T_na+T_nb+T_nc+T_nd+........T_nn}{T_F * T_n} \end{equation} RB-Bayes classifier have minimum error rate as compared to other algorithms. All factors are taken into consideration \section{Usage and Comparison of proposed methods} Python is used for implementation of methodology. We build RB-based algorithm based on baye’s theorem. Preprocessing steps done before applying proposed algorithm is shown in Figure 2.After we have dataset on which we want to implement this algorithm. We need to perform some preprocessing steps.Seperate tuple from label on which we want to do prediction. Single row represent tuple. We check it on small datasets as shown in Figure 3.compare with naïve baye’s also. Dataset contains text data also. So need to convert this data into numeric form. Some categorical variables consist of more than two values. So after converting dataset into numeric, dataset consist of values more than 0 and 1 also depending on category. We need to do dummy encoding of those variables which consists more than two values. \begin{figure} \centering \includegraphics[width=0.5\textwidth]{prediction.jpg} \caption{Methodology used for prediction.} \end{figure} \par Our dataset is as great blend of categorical and continuous qualities and fills in as a helpful case that is moderately simple to understand. Thus, the examiner is looked with the test of making sense of how to transform these content traits into numerical qualities for encourage processing. Label encoding has the preferred standpoint that it is direct however it has the drawback that the numeric qualities can be "confounded" by the calculations. A typical elective approach is called one hot encoding. In spite of the diverse names, the fundamental methodology is to change over every classification esteem into another segment and appoints a 1 or 0 (True/False) esteem to the segment. \end{enumerate} \begin{table} \caption{Dataset} \begin{tabular}{|l|l|l|l|l|} \hline Age & Income & Student & Credit rating & Buys Computer \\ \hline youth & high & no & fair & no \\ \hline youth & high & no & excellent & no \\ \hline middle aged & high & no & fair & yes \\ \hline senior & medium & no & fair & yes \\ \hline senior & low & yes & fair & yes \\ \hline senior & low & yes & excellent & no \\ \hline middle aged & low & yes & excellent & yes \\ \hline youth & medium & no & fair & no \\ \hline youth & low & yes & fair & yes \\ \hline senior & medium & yes & fair & yes \\ \hline youth & medium & yes & excellent & yes \\ \hline middle aged & medium & no & excellent & yes \\ \hline middle aged & high & yes & fair & yes \\ \hline senior & medium & no & excellent & no \\ \hline \hline \end{tabular} \end{table} \par As we know python is based on mathematical equations, so all dataset must in binary form. But our dataset contains text data. First step is to convert all dataset into numeric. We import class Label Encoder from sklearn to change data from text into numeric. Age and income contains three categories. Under age categories youth converted into 2, middle-aged converted into 0 and last senior converted into 1.Under income categories high is converted into 0,medium converted into 2 and low is converted into 1.OneHotEncoder class will convert this data from numeric to binary because python understand only binary data. Dummy encoding will be generated by OneHotEncoder class as shown in Table 2.Data has been converted into binary form. \begin{table} \caption{Dummy Encoding} \begin{tabular}{|l|l|l|l|l|l|l|l|} \hline Age & middle aged & Senior & Youth & Income & High & Low & Medium \\ \hline youth & 0 & 0 & 1 & high & 1 & 0 & 0 \\ \hline youth & 0 & 0 & 1 & high & 1 & 0 & 0 \\ \hline middle aged & 1 & 0 & 0 & high & 1 & 0 & 0 \\ \hline senior & 0 & 1 & 0 & medium & 0 & 0 & 1 \\ \hline senior & 0 & 1 & 0 & low & 0 & 1 & 0 \\ \hline senior & 0 & 1 & 0 & low & 0 & 1 & 0 \\ \hline middle aged & 1 & 0 & 0 & low & 0 & 1 & 0 \\ \hline youth & 0 & 0 & 1 & medium & 0 & 0 & 1 \\ \hline youth & 0 & 0 & 1 & low & 0 & 1 & 0 \\ \hline senior & 0 & 1 & 0 & medium & 0 & 0 & 1 \\ \hline youth & 0 & 0 & 1 & medium & 0 & 0 & 1 \\ \hline middle aged & 1 & 0 & 0 & medium & 0 & 0 & 1 \\ \hline middle aged & 1 & 0 & 0 & high & 1 & 0 & 0 \\ \hline senior & 0 & 1 & 0 & medium & 0 & 0 & 1 \\ \hline \end{tabular} \end{table} \par We compare the result with naïve Bayes. Suppose we wish to predict value for below tuple whether this tuple will purchase computer or not.\\ X = (age=youth ,income=medium , student=yes,credit rating=fair) Using RB-Bayes algorithm, we are going to predict the possibility for above tuple. \begin{equation} Mean_{yes} = \frac{T_y}{Total_SampleSet} \end{equation} \begin{equation} Mean_{no} = \frac{T_n}{Total_SampleSet} \end{equation} Total yes consist of all records that purchases computer and Total no consist of those who do not purchase computer. So, we are calculating mean. mean\textsubscript {yes}=0.64 and mean\textsubscript {no}=0.36 and total number of samples =14 \par After calculating mean, now we take summation of all factors who said yes to purchase computer and similarly calculating for those who said no for purchasing computer \\ \begin{center} F\textsubscript y=Number of factors * Total\textsubscript{yes} F\textsubscript n=Number of factors * Total\textsubscript{no} \end{center} \begin{equation} \begin{aligned} P_yF =\frac{T_ya+T_yb+T_yc+T_yd }{F_y} \\ P_nF =\frac{T_na+T_nb+T_nc+T_nd }{F_n} \end{aligned} \end{equation} \par T\textsubscript ya , T\textsubscript yb , T\textsubscript yc and T\textsubscript yd are total number of factors where T\textsubscript ya is one and even value for label is also one. Where both of the condition is true that gives value for these factors. Total\textsubscript {yes} count number of ‘yes’ from total number of samples.Total\textsubscript {no} count number of no from total number of samples.. To calculate value for probability for yes or no we multiply value with mean\textsubscript yes and mean\textsubscript no respectively after calculating summation of P\textsubscript yF and P\textsubscript nF. \\ P\textsubscript yF=18 \\ P\textsubscript nF=8 \\ \begin{center} P\textsubscript{yes}=P\textsubscript yF * mean\textsubscript {yes} \\ P\textsubscript {no}=P\textsubscript nF * mean\textsubscript {no} \end{center} P\textsubscript {yes}=0.32 \\ P\textsubscript {no}=0.14 \\ Compare values for probability of yes and probability of no to find greater value. Greater value will decide whether particular tuple will purchase a computer or not. Value for P\textsubscript yF is greater than P\textsubscript nF. So we can predict that this tuple will purchase computer. Calculating Accuracy Our algorithm removes the problem of zero probability and also improves accuracy. To find this we divide our data into training and test set. We set test size = 0.37.We test same dataset using naïve bayes and RB-bayes algorithm also to check accuracy. After calculating value for probability of yes i.e. P\textsubscript {yes} and probability of no i.e. P\textsubscript {no}, we compare these two values and highest value does prediction. We test the accuracy score in python using class accuracy score. \begin{center} from sklearn.metrics import accuracy\_score \\ print('Accuracy score:', accuracy\_score(y\_test,y\_pred)) \end{center} Using RB-Bayes algorithm ,value of accuracy =83.3% Using Naïve bayes algorithm, value of accuracy = 50% Reason is in Naïve Bayes when we end up with probability of zero ,we lose effect for other factors also. Although we use Laplace correction that each one value in each account but in actual value is zero. IN RB-Bayes, we remove possibility of zero. \par In machine learning, Support vector machines (SVMs, additionally bolster vector systems) are directed learning models with related learning calculations that break down information utilized for grouping and relapse examination. We apply same dataset on SVM also for comparison.SVM methodology is implemented in Rapid miner as shown in Figure 5. \begin{figure} \centering \includegraphics[width=0.5\textwidth]{SVM.jpg} \caption{Implementation with SVM.} \end{figure} There are various bibliography styles available. You can select the style of your choice in the preamble of this document. These styles are Elsevier styles based on standard styles like Harvard and Vancouver. Please use Bib\TeX\ to generate your bibliography and include DOIs whenever available. \par Nominal to numerical operator is used to convert text data into numerical. Before applying SVM, it is require to convert data into numerical form. Performance operator is used to test the accuracy of model. Confusion matrix is generated as shown in Figure 6. This Operator is utilized to measurably assess the qualities and shortcomings of a double order, after a prepared model has been connected to named data. A paired arrangement makes forecasts where the result has two conceivable qualities: call them positive and negative. In addition, the forecast for every Example might be correct or wrong, prompting \\ TP - the quantity of "genuine positives", positive Examples that have been accurately distinguished. \\ FP - the quantity of "false positives", negative Examples that have been inaccurately recognized. \\ FN - the quantity of "false negatives", positive Examples that have been inaccurately recognized. \\ TN - the quantity of "genuine negatives", negative Examples that have been accurately recognized.\\ \begin{figure} \centering \includegraphics[width=0.5\textwidth]{CM.jpg} \caption{Confusion Matrix} \end{figure} \begin{figure} \centering \includegraphics[width=0.5\textwidth]{SVMACC.jpg} \caption{Accuracy using SVM} \end{figure} \begin{table} \centering \caption{Accuracy and comparisons of algorithms} \begin{tabular}{|l|l|}\hline Algorithm & Accuracy \\ \hline RB-Bayes Algorithm & 83.3\% \\ \hline Naive Bayes Algorithm & 50\% \\ \hline Support Vector machine Algorithm & 85.71\% \\ \hline \end{tabular} \end{table} 83.3 is not bad accuracy of RB-Bayes algorithm as shown in table 3. So we can characterize exactness measures of model as a component of the check accurately anticipated records. \section{Conclusion and future work} In this paper, we study the supervised techniques which allow doing prediction based on training data. Naive bayes algorithm for data mining has been reviewed and a new approach is proposed. It is important to stress that the proposed algorithm consider all factors even if probability of likelihood is zero. Apart from the existing supervised techniques, this model may also be of interest in market where manufacturers or seller wants to know why their sale is up or down. On what factors they need to give importance or work. They can improve sales performance. We can know what the factors affect the buying decision of customer are. Tests are directed to confirm this calculation for small datasets and promising outcomes are acquired. In future, this idea of amalgamation of bunching and characterization can be connected over enormous information influencing utilization of guide to diminish method to deal with vast databases. \section{Bibliography styles} %\bibliographystyle{IEEE} \begin{thebibliography}{00} \bibitem{1} Xu, L.E.I., Jiang, C., Wang, J., 2014. Information Security in Big Data : Privacy and Data Mining 1149–1176. \bibitem{2}Turney, P.D., 2002. Thumbs up or thumbs down? Semantic Orientation applied to Unsupervised Classification of Reviews. Proc. 40th Annu. Meet. Assoc. Comput. Linguist. 417–424. https://doi.org/10.3115/1073083.1073153. \bibitem{3}Pang, B., Lee, L., Vaithyanathan, S., 2002. Thumbs up?: sentiment classification using machine learning techniques. Proc. Conf. Empir. Methods Nat. Lang. Process. 79–86. https://doi.org/10.3115/1118693.1118704. \bibitem{4}Mihalcea, R., Banea, C., Wiebe, J., 2007. Learning multilingual subjective language via cross-lingual projections. Proc. 45th Annu. Meet. Assoc. Comput. Linguist. 976–983. https://doi.org/citeulike-article-id:3270776. \bibitem{5}Su, F., Su, F., Markert, K., Markert, K., 2008. From words to senses: a case study of subjectivity recognition. Proc. 22nd Int. Conf. Comput. Linguist. 1 825–832. \bibitem{6}Pang, B., Lee, L., 2004. A Sentimental Education: Summarization Based on Minimum Cuts. https://doi.org/10.3115/1218955.1218990. \bibitem{7}Pang, B., Lee, L., 2006. Opinion Mining and Sentiment Analysis. Found. Trends® InformatioPang, B., Lee, L. (2006). Opin. Min. Sentim. Anal. Found. Trends® Inf. Retrieval, 1(2), 91–231. doi10.1561/1500000001n Retr. 1, 91–231. https://doi.org/10.1561/1500000001. \bibitem{8}Lim, E.-P., Nguyen, V.-A., Jindal, N., Liu, B., Lauw, H.W., 2010. Detecting product review spammers using rating behaviors. Proc. 19th ACM Int. Conf. Inf. Knowl. Manag. - CIKM ’10 939. https://doi.org/10.1145/1871437.1871557. \bibitem{9}Jindal, N., Liu, B., 2008. Opinion spam and analysis. Proc. Int. Conf. Web search web data Min. WSDM 08 219. https://doi.org/10.1145/1341531.1341560. \bibitem{10}Jindal, N., Liu, B., Street, S.M., 2007. Review Spam Detection 1189–1190. \bibitem{11}Liu, B., Street, S.M., 2005. Opinion Observer : Analyzing and Comparing Opinions on the Web. Proc. 14th Int. Conf. World Wide Web 342–351. https://doi.org/10.1145/1060745.1060797. \bibitem{12}Morinaga, S., Yamanishi, K., Tateishi, K., Fukushima, T., 2002. Mining product reputations on the Web. Proc. eighth ACM SIGKDD Int. Conf. Knowl. Discov. data Min. - KDD ’02 341. https://doi.org/10.1145/775094.775098. \bibitem{13}Li, F., Han, C., Huang, M., Zhu, X., Xia, Y.-J., Zhang, S., Yu, H., 2010. Structure-aware Review Mining and Summarization. Proc. 23rd Int. Conf. Comput. Linguist. 653–661. \bibitem{14}Jindal, N., Liu, B., 2006. Identifying comparative sentences in text documents. Proc. 29th Annu. Int. ACM SIGIR Conf. Res. Dev. Inf. Retr. - SIGIR ’06 244. https://doi.org/10.1145/1148170.1148215. \bibitem{15}Kim, H.D., Zhai, C., 2009. Generating comparative summaries of contradictory opinions in text. Proceeding 18th ACM Conf. Inf. Knowl. Manag. - CIKM ’09 385. https://doi.org/10.1145/1645953.1646004. \bibitem{16}Hu, M., Liu, B., 2004. Mining and summarizing customer reviews. Proc. 2004 ACM SIGKDD Int. Conf. Knowl. Discov. data Min. KDD 04 4, 168. https://doi.org/10.1145/1014052.1014073. \bibitem{17}Jo, Y., Oh, A.H., 2011. Aspect and sentiment unification model for online review analysis. Proc. fourth ACM Int. Conf. Web search data Min. - WSDM ’11 815. https://doi.org/10.1145/1935826.1935932. \bibitem{18} Wu, Y., Zhang, Q., Huang, X.X., Wu, L., 2009. Phrase Dependency Parsing for Opinion Mining. Proc. 2009 Conf. Empir. Methods Nat. Lang. Process. Vol. 3 EMNLP 09 1533–1541. https://doi.org/10.3115/1699648.1699700 \bibitem{19}Snyder, B., Barzilay, R., 2005. Multiple Aspect Ranking using the Good Grief Algorithm. \bibitem{20}Titov, I., McDonald, R., 2008. A joint model of text and aspect ratings for sentiment summarization. Proc. ACL08 HLT 51, 308–316. https://doi.org/10.1039/b003067h. \bibitem{21}Pham, D.H., Le, A.C., Le, T.K.C., 2015. A least square based model for rating aspects and identifying important aspects on review text data. Proc. 2015 2nd Natl. Found. Sci. Technol. Dev. Conf. Inf. Comput. Sci. NICS 2015 265–270. https://doi.org/10.1109/NICS.2015.7302204. \bibitem{22}Pham, D.H., Le, A.C., Le, T.K.T.and others 2016, Determining Aspect Ratings and Aspect Weights from Textual Reviews by Using Neural Network with Paragraph Vector Model, in: Nguyen, H.T., Snasel, V. (Eds.), International Conference on Computational Social Networks. Springer International Publishing, Vietnam, pp. 309-320. https://doi.org/10.1007/978-3-319-42345. \bibitem{23}Zha, Z.J., Yu, J., Tang, J., Wang, M., Chua, T.S., 2014. Product aspect ranking and its applications. IEEE Trans. Knowl. Data Eng. 26, 1211–1224. https://doi.org/10.1109/TKDE.2013.136 \bibitem{24}Pham, D.-H., Le, A.-C., 2016. A Neural Network based Model for Determining Overall Aspect Weights in Opinion Mining and Sentiment Analysis. Indian J. Sci. Technol. 9, 1–6. https://doi.org/10.17485/ijst/2016/v9i18/93164 \bibitem{25}Hartog, J., van Zanten, H., 2018. Nonparametric Bayesian label prediction on a graph. Comput. Stat. Data Anal. 120, 111–131. https://doi.org/10.1016/j.csda.2017.11.008 \bibitem{26}Muthulakshmi, S., Dash, C.S., Prabaharan, S.R.S., 2018. Memristor augmented approximate adders and subtractors for image processing applications: An approach. AEU - Int. J. Electron. Commun. 91, 91–102. https://doi.org/10.1016/j.aeue.2018.05.003 \bibitem{27}Abdulrazzaq, M.M., Yaseen, I.F.T., Noah, S.A., Fadhil, M.A., 2018. Multi-Level of Feature Extraction and Classification for X-Ray Medical Image 10, 154–167. https://doi.org/10.11591/ijeecs.v10.i1.pp154-167 \bibitem{28}Hastarita Rachman, F., Sarno, R., Fatichah, C., 2018. Music Emotion Classification based on Lyrics-Audio using Corpus based Emotion. Int. J. Electr. Comput. Eng. 8, 1720–1730. https://doi.org/10.11591/ijece.v8i3.pp1720-1730 \end{thebibliography} \end{document}