Social media provides data that can be used for text classification, but these social media do not have a rating system. Comments taken from social media such as Twitter, does not provide a rating system which can help to classify comments based on their rating score. The goal of this study is to build a comment classification model using Word2vec and SVM classifier that can classify comments based on a rating scale from 1-5. The training data will be taken from user comments from play.google.com website in which each comment already has a rating. The purpose of the model is to classify comments from social media about mobile network applications. Comment classification is performed in order to help these businesses, as well as the users, to know the overall satisfaction of users who use these applications. The best F1 score obtained from this research using class elimination and stop word removal is 0.795.

This study will classify the text based on the rating of the provider application on the Google Play Store. This research is classification of user comments using Word2vec and the deep learning algorithm in this case is Long Short Term Memory (LSTM) based on the rating given with a rating scale of 1-5 with a detailed rating 1 is the lowest and rating 5 is the highest data and a rating scale of 1-3 with a detailed rating, 1 as a negative is a combination of ratings 1 and 2, rating 2 as a neutral is rating 3, and rating 3 as a positive is a combination of ratings 4 and 5 to get sentiment from users using SMOTE oversampling to handle the imbalance data. The data used are 16369 data. The training data and the testing data will be taken from user comments MyTelkomsel’s application from the play.google.com site where each comment has a rating in Indonesian Language. This review data will be very useful for companies to make business decisions. This data can be obtained from social media, but social media does not provide a rating feature for every user comment. This research goal is that data from social media such as Twitter or Facebook can also quickly find out the total of the user satisfaction based from the rating from the comment given. The best f1 scores and precisions obtained using 5 classes with LSTM and SMOTE were 0.62 and 0.70 and the best f1 scores and precisions obtained using 3 classes with LSTM and SMOTE were 0.86 and 0.87