Effect of Various Factors in Context of Feature Selection on Opinion Spam Detection

With the growing popularity of online reviews, spammers often target specific products or services with the aim to mislead consumers in their purchase decisions. This has opened doors for researchers to study the problem of opinion spam detection. Till date, many effective and efficient solutions have been proposed in this regard using various types of features. However, most of the feature engineering tasks extract thousands of features, which may lead to degrade the performance and increase computation cost involved in many machine learning algorithms. Feature selection methods can greatly improve classification performance along with the reduction in computation cost of model training. In this paper, we investigate the effect of different feature selection techniques on opinion spam detection. For the same, various feature selection methods (filter-based and model-based) with varying number of features have been employed to train four different classification models. In addition, three well-known review datasets from different domains (hotel, doctor and restaurant) and four different types of features, viz., unigram, bigram, part-of-speech frequency count and word embedding, have been used to examine the impact of different factors responsible to improve the performance in opinion spam domain. Our experimental results demonstrate how different factors affect classification performance and cost, which is statistically validated by using Analysis of Variance test.