An improved semi-supervised dimensionality reduction using feature weighting: Application to sentiment analysis

• Published in: Expert systems with applications 2018-11, Vol.109, p.49-65
• Main Author: Kim, Kyoungok
• Format: Article
• Language: English
• Subjects: Algorithms; Benchmarks; Classification; Cost analysis; Data mining; Dimensional analysis; Feature extraction; Feature weighting; Mapping; Natural language processing; Natural language processing (NLP); Reduction; Semi-supervised dimensionality reduction; Sentiment analysis; Weighting
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2018.05.023

•A semi-supervised feature extraction combines with feature weighting is proposed.•Feature weighting considers both co-occurrence of terms and label of documents.•The polarity scores defined in SentiWordNet are reflected in the feature weights.•Six datasets are used to validate the enhanced performance of the proposed method. Analyzing a large number of documents for sentiment analysis entails huge complexity and cost. To alleviate this burden, dimensionality reduction has been applied to documents as a preprocessing step. Among dimensionality reduction algorithms, compared with feature selection, feature extraction can reduce information loss and achieve a higher discriminating power in sentiment classification. However, feature extraction suffers from lack of interpretability and many nonlinear extraction methods, which generally outperform linear methods, are not applicable for sentiment classification because of the characteristics that only provide corresponding low-dimensional coordinates without mapping. Therefore, this research proposes an improved semi-supervised dimensionality reduction framework that simultaneously preserves the advantages of feature extraction and addresses the drawbacks for sentiment classification. The proposed framework is mainly based on linear feature extraction providing mapping and feature weighting is applied before feature extraction. Feature weighting and extraction are conducted in a semi-supervised manner so that both label information and structural information of data can be considered. The superiority of both feature weighting and feature extraction was verified by conducting extensive experiments in six benchmark datasets.