Abstract:
Customer churn prediction is one of the major issues that any business needs to address in order to survive. In telecom services, churn prediction is not only challenging but also very expensive. There are already several products with different methods to track customer churn and many such analytics tools that can be used to provide this information. Sentiment analysis is an emerging field that frames data around subjective opinions, attitudes and emotions expressed by people about a topic. In order to implement sentiment analysis, it is necessary to first collect data that can be used for sentiment classification. The theory is that when customers are happy with their service providers, they are more likely to stick with them than when they experience dissatisfaction or frustrations. In this research over 2000 reviews based on social media data were analyzed to predict whether the reviewer will churn or not based on the sentiment. The dataset was preprocessed first and two major feature extraction techniques, CountVectorizer and TF-IDF vectorizer were applied. Finally state of the art machine learning models were created for each feature extraction technique to confirm the best method with best ML algorithm. The result indicates that in Count-Vector technique the Bagging classifier surpassed other models with a 98% accuracy while Multi-Layer Perceptron scores 99% accuracy for the TF-IDF feature extraction technique.
