Developing an LSTM-based Classification Model of IndiHome Customer Feedbacks

The development of companies in the service sector is inseparable from problems with the customer service process. The latest trends show that customers are more comfortable using social media and chat services as feedback delivery media compared to sending emails.In this study, we classify user feedback from an Internet Service Provider (ISP) in Indonesia named IndiHome. In this study, there are two test scenarios carried out to classify, the first scenario uses Short Term Memory combined with Word Embedding, and the second scenario uses Naive Bayes combined with TF-IDF. based on the test results, obtained accuracy on the LSTM method combined with Word Embedding with an f1 score of 87.98 % and accuracy obtained at Naive Bayes combined with TF-IDF with an f1 score of 76.77 %. From the tests that have been done, different results are obtained and have a big difference, which is 11.21%. The difference in the results of this classification is influenced by several factors, namely the first factor is the extraction of features used and the second is the data used. In feature selection, TF-IDF is more ideal for data that has a large or long document size because the representation seen from the whole document, in contrast to Word Embedding that is not affected by the data size because the representation is seen from words. so this is what causes the method of using LSTM combined with Word Embedding to produce higher results.