An Experimental Study of Supervised Sentiment Analysis Using Gaussian Naïve Bayes Unggul Widodo Wijayanto Informatics Department, Faculty of Technology Information and Communication Institut Teknologi Sepuluh Nopember Surabaya, Indonesia unggul15@mhs.if.its.ac.id Riyanarto Sarno Informatics Department, Faculty of Technology Information and Communication Institut Teknologi Sepuluh Nopember Surabaya, Indonesia riyanarto@if.its.ac.id Abstract— In this millennial generation everyone using technology at higher rates than people from other generations. It means the millennial generation is aware of evolving technology. Many companies are taking chances by receiving customer reviews through applications. In this study, we use customer reviews from Yelp (foods), IMDb (movies) and Amazon (products). The reviews received by the company are numerous. Product management does not have much time to read customer reviews one by one. So, to speed up the reading of customer reviews we were using sentiment analysis. There are many methods that used in sentiment analysis such as supervised sentiment analysis. We used TF-IDF to convert word to features implements the supervised method. Performance of the supervised method depends on the data training quality. So, to improve the accuracy of the results by improving data training quality. The methods used to improve the data training quality in this paper using CHI2 Features Selection and Stopwords. In this study, we use K-folds Cross-validation to get valid results. This study proves the use of Context-based Stopwords can improve the results. Context-based Stopwords enrich the number of Stopwords that removing bias features. Keywords—supervised, sentiment analysis, data training, text mining, cross-validation, features selection, stopwords. I. INTRODUCTION In this era, people can do anything on the internet. The Internet World Stats state that there are 4.1 billion internet users in the world on December 31, 2017 [1]. Based on internet world stats on December 31, 2017 internet user growth 1,052% for 18 years [1]. Besides using the internet to search for information they can also review. Internet users often review food, movies, and products. There are many websites that provide user reviews, like Yelp.com, IMDb.com, and Amazon.com. Product reviews can help a company's decision-making. For the example is decision making in a cinema company. They can predict a good movie that will be of interest to many people using reviews from IMDb.com. When good reviews received massively in the first week. Cinemas can open more theaters. If the opposite happens, the cinema may reduce the duration of the theater. We can classify good and bad review by using sentiment analysis. Some applications from Natural Language Processing (NLP) iFs opinion mining and sentiment analysis, text analysis to find out and extract the meaning of the information obtained from the review document, blogs or any source materials. And positive or negative can be the pole of a sentence [2]. Example, “This is definitely a cult classic well worth viewing and sharing with others.” it means the audience has a positive sentiment to the movie. Another example, “Otherwise, don't even waste your time on this.” It means the audience has a negative sentiment when watching the movie. Sentiment analysis or usually called opinion mining had been studied many times. With the development of connectivity technology that provides many ways to interpret and process the user opinion. Some use machine learning methods such as Naïve Bayes, Maximum Entropy and Support Vector Machines [3]. In this study, we use Gaussian Naive Bayes. This experimental study used supervised methods. Supervised methods have a good accuracy, but it depends on the data train that gives in the preprocessing. The methods used to improve the quality of data training in this paper using CHI2 features selection and Stopwords. The CHI2 algorithm is based on the χ² statistic to select the minimum number of attributes. Stopwords by definition are meaningless words that have low discrimination power (Lo et al., 2005) [4]. Stopwords are deleting meaningless words to avoid ambiguous results. In this paper, we compare the correlations between the 3 datasets with the number of features selected. In addition, we also compared the effect of using General Stopwords and Context-based Stopwords. General Stopword is obtained from the function on TF-IDF. While Context-based Stopwords is obtained from each dataset. This paper is organized as follows: Section II contains similar research has ever done; Section III explains the methods used in this study; Section IV consists of an explanation of the results; Section V consists of conclusions and future research. II. RELATED WORKS Research on sentiment analysis by classifying customer reviews using machine learning or lexicon-based has been widely practiced. S. Rani and P. Kumar [5] conducting research on sentiment analysis system which aims to improve teaching and learning. This research uses natural language processing and machine learning to analyze student feedbacks. 2018 International Seminar on Application for Technology of Information and Communication (iSemantic) 476