Dimensionality reduction or the optimal selection of features is a challenging task due to large search space. Currently, many research has been performed in this domain to improve the accuracy as well as to minimize the computational complexity. Particle Swarm Optimization (PSO) based feature selection approach seems very promising and has been extensively used for this work. In this paper, a Threshold Controlled Binary Particle Swarm Optimization (TC-BPSO) along with Multi-Class Support Vector Machine (MC-SVM) is proposed and compared with Conventional Binary Particle Swarm Optimization (C-BPSO). TC-BPSO is used for the selection of features while MC-SVM is used to calculate the classification accuracy. 70% of the data is used to train the MC-SVM model while the test has been performed on rest 30% data to calculate the accuracy. Proposed approach is tested on ten different datasets having varying difficulties such as some datasets having large number of features while some have small, some have just two classes while some have many classes, some datasets having small number of instances while some have large number of instances and the results obtained on these datasets are compared with some of the existing methods. Experiments show that the obtained results are very promising and achieved the best accuracy in minimum possible features. Proposed approach outperforms C-BPSO in all contexts on most of the datasets and 3-4 times computationally faster. It also outperforms in all context when compared with the existing work and 5-8 times computationally faster.

Sentiment Classification is a special task of Sentiments Analysis in which a text document is assigned into some category like positive, negative, and neutral on the basis of some subjective information contained in documents. This subjective information called as sentiment features are highly responsible for efficient sentiment classification. Thus, Feature extraction is essentially an important task for sentiment classification at any level. This study explores most relevant and crucial features for sentiment classification and groups them into seven categories, named as, Basic features, Seed word features, TF-IDF, Punctuation based features, Sentence based features, N-grams, and POS lexicons. This paper proposes two new sentence based features which are helpful in assigning the overall sentiment of contrastive sentences and on the basis of proposed features; two algorithms are developed to find the sentiment of contrastive sentences. The dataset of TripAdvisor is used to evaluate our proposed features. Obtained results are compared with several state-of-the-art studies using various features on the same dataset and achieve superior performance.