Reducing Feature Set Explosion to Facilitate Real-World Review Spam Detection
Mike Crawford, Taghi M. Khoshgoftaar, Joseph D. Prusa · The Florida AI Research Society · 2016
Online reviews are quickly becoming one of the most important sources of information for consumers on various products and services. With their increased importance, there exists an increased opportunity for spammers or unethical business owners to create false reviews in order to artificially promote their goods and services or smear those of their competitors. In response to this growing problem, there have been many studies on the most effective ways of detecting review spam using various machine learning algorithms. One common thread in most of these studies is the conversion of reviews to word vectors, which can potentially result in hundreds of thousands of features. However, there has been little study on reducing the feature subset size to a manageable number or how best to do so. In this paper, we consider two distinct methods of reducing feature subset size in the review spam domain. The methods include filter-based feature rankers and word-frequency based feature selection. We show that there is not a one size fits all approach to feature selection, and the best way to reduce the feature subset size is dependent upon both the classifier being used and the feature subset size desired. It was also observed that the feature subset size had significant influence on which feature selection method is used.