Flower Pollination Algorithm for Feature Selection in Tweets Sentiment Analysis

Text-based social media platforms have developed into important components for communication between customers and businesses. Users can easily state their thoughts and evaluations about products or services on social media. Machine learning algorithms have been hailed as one of the most efficient approaches for sentiment analysis in recent years. However, as the number of online reviews increases, the dimensionality of text data increases significantly. Due to the dimensionality issue, the performance of machine learning methods has been degraded. However, traditional feature selection methods select attributes based on their popularity, which typically does not improve classification performance. This work presents a population-based metaheuristic for feature selection algorithms named Flower Pollination Algorithms (FPA) because of their propensity to accept less optimum solutions and avoid getting caught in local optimum solutions. The study analyses tweets from Kaggle first with the usual Term Frequency-Inverse Document Frequency statistical weighting filter and then with the FPA. Four baseline classifiers are used to train the features: Naive Bayes (NB), Decision Tree (DT), Support Vector Machine (SVM), and k-Nearest Neighbor (kNN). The results demonstrate that the FPA outperforms alternative feature subset selection algorithms. For the FPA, an average improvement in accuracy of 2.7% is seen. The SVM achieves a better accuracy of 98.99%.