Comparative Study on Swarm Based Algorithms for Feature Reduction in Twitter Sentiment Analysis on Figurative Language

Abstract

In this paper, we deal with the issue of feature selection by comparing different approaches based on Gravitational Search Algorithm, Particle Swarm Optimisation and Genetic Algorithm. The comparison is drawn on the parameters of feature reduction percentage, accuracy and time taken. The optimization is performed with the following supervised predictive models - Multinomial Naive Bayes, Support Vector Machine, Decision Tree, K-Nearest Neighbour and Multilayer Perceptron. Datasets were acquired from SemEval 2015 task on the sentiment analysis of figurative language on Twitter (Task 11), which provided a set of 5198 tweets scored with finely grained sentiment score (−5 to +5 including 0). 11538 features were generated from the datasets and the experiments performed have been successful in reducing an average of 55% of the features, without any decline in the accuracy.