Semantic similarity-aware feature selection and redundancy removal for text classification using joint mutual information

Abstract

The high dimensionality of text data is a challenging issue that requires efficient methods to reduce vector space and improve classification accuracy. Existing filter-based methods fail to address the redundancy issue, resulting in the selection of irrelevant and redundant features. Information theory-based methods effectively solve this problem but are not practical for large amounts of data due to their high time complexity. The proposed method, termed semantic similarity-aware feature selection and redundancy removal (SS-FSRR), employs joint mutual information between the pairs of semantically related terms and the class label to capture redundant features. It is predicated on the assumption that semantically related terms imply potentially redundant ones, which can significantly reduce execution time by avoiding sequential search strategies. In this work, we use Word2Vec’s CBOW model to obtain semantic similarity between terms. The efficiency of the SS-FSRR is compared to six state-of-the-art competitive selection methods for categorical data using two traditional classifiers (SVM and NB) and a robust deep learning model (LSTM) on seven datasets with 10-fold cross-validation, where experimental results show that the SS-FSRR outperforms the other methods on most tested datasets with high stability as measured by the Jaccard’s Index.