Sentiment analysis has been very effective to obtain useful insights from texts, such as the polarity – positive or negative, or the category based on emotional tendency or contents, for academic, political or financial purpose. News, a structured and stylistic type of articles, as “factual writing”, different writers have bias in selecting the source, that’s why sentiment analysis can be applied in news. Feature selection has been a popular technique to find feature subset from feature sparse matrix transformed from the original texts to enhance the accuracy of the classification model. Metaheuristics has been applied as learning methods to improve the flexibility and performance of the prediction model. Feature selection is a common used approach combined with different learning methods to reduce the dimension of sparse dataset. The sparsity of the feature matrix is usually very high, more than 95%. Swarm search-based feature selection is a fusion model between the swarm search methods and feature selection. In this paper, a new feature selection method called Optimized Swarm Search-based Feature Selection (OS-FS) is applied, Best First is chosen as the standard baseline, while Evolutionary Search and PSO Search are chosen as swarm search based algorithm. The OS-FS model is optimized by Clustering-by-Coefficient-of-Variation (CCV) is as comparison. Clustering-by-Coefficient-of-Variation (CCV) method selects a feature subset as the starting point for other learning method. Compared with original version of swarm FS, CV-based FS achieves higher accuracy without costing much time, has potential to be modified for parallel processing. The proposed model is experimented via three datasets, a binary class of sentiments, positive or negative, or category of sentiments, financial, sports, and politics.