Feature Selection based Sentiment Analysis on US Airline Twitter Data

Abstract

Review document emotions are classified using sentiment analysis. Researchers grade features to remove non-informative and noisy attributes with low grades to improve classification accuracy. This paper utilizes six different NLP models to predict user sentiments based on Twitter reviews about airlines, using the Twitter US Airline Sentiment dataset. The best-performing models from both machine learning (K-nearest neighbor, Random forest, and Multinomial Naive Bayes) and deep learning (Artificial Neural Networks, LSTM, and Bidirectional LSTM with Glove embeddings) were implemented through Anaconda and Google Colab platforms. This paper introduces a new type of feature dimensionality technique termed "inquiry extension grade (IEG)," inspired by the inquiry extension term weighting technique. Additionally, we modified the traditional TF-IDF method, referred to as "improved TF-IIDF (IFFIDF)," specifically tailored for processing unbalanced text collections. To assess the effectiveness of the proposed methods, a series of simulations were conducted. The results indicate that the combination of IEG-ITFIDF Vectorization and Bi-LSTM with Glove embeddings yielded the best accuracy of 94.26% in sentiment classification for the Twitter US Airline Sentiment dataset.