Optimizing the Performance of Sentiment Analysis on Social Media Data Through Feature Selection in Preprocessing Techniques

The digital era directly builds user needs in communicating using social media, the data generated by social media has a large volume so it needs to be analyzed with a preprocessing process. The problems that exist in social media data contain a lot of repeated words with non-standard sentence structures and still contain lots of irregular punctuation making it difficult to determine which features are considered important in a sentence to carry out a sentiment analysis process in identifying positive, negative, and neutral sentences. So it is necessary to select features using preprocessing techniques using Recursive Feature Elimination (RFE). This study produces a comparison that shows that the selection of features and combinations in using features results in the accuracy of each algorithm increasing when compared to using one feature. The results showed that the CNN algorithm has an accuracy value of 0.95, the LSTM algorithm is 0.94 and the RNN algorithm is 0.93. when compared to the use of one feature, it increases by 0.09 for the CNN algorithm, increases by 0.07 for the LSTM algorithm, and increases by 0.05 for the RNN algorithm.