Particle Swarm Optimized Ensemble Learning for Enhanced Predictive Sentiment Accuracy of Tweets

Abstract

The mounting global interest of users in social media portals has reinforced research in knowledge discovery domains to mine useful information from the publicly available user-generated big data. Uncertainty is often linked with the online content, mostly owing to the diverse, noisy, or unstructured data which might be imprecise or vague. Determining appropriate features that could yield enhanced sentiment predictive accuracy becomes a tedious and an arduous task which motivates for automating the sentiment classification predictive task. The empirical study is investigated on the tweets fetched from two standard Twitter datasets namely SemEval 2016 & SemEval 2017. Consequently, we propose the use of swarm based feature selection for enhancing classifier efficiency. The proposed feature selection using PSO outperforms baseline ensemble learning algorithm trained using conventional tf-idf. An average 8.5% improvement in accuracy with 33% reduction in feature set is obtained by implementing particle swarm optimization.