Opinion Spam Detection with Feature Selection and PU Learning

Abstract

In current times, it has gotten extremely vital for e-business organizations to enable their end clients to write opinion about the items/services that they have used. Such reviews provide vital sources of information on these products/services. This data is used by the future potential clients before choosing buy of new items or services. These opinions or reviews are also exploited by marketers to find out the drawbacks of their own products/services and alternatively to find the very important information related to their competitor’s products/services. Unfortunately this significant usefulness of opinions has also raised the problem for spam, which contains forged positive or spiteful negative opinions.

A recently proposed opinion spam detection method which is based on n-gram techniques is extended by means of feature selection and different representation of the opinions. The problem is modelled as the classification problem and naïve-Bayes classifier and least-square support vector machine (LS-SVM) are used on three different representations (Boolean, bag-of-words and term frequency inverse document frequency (TF-IDF)) of the opinions. All the experiments are carried out on widely used gold-standard dataset and got encouraging result.

Also proposed a method to learn a classifier for the task of opinion spam detection in the presence of only small number of positive opinions (e.g. spam opinions). This method is inspired by a methodology named learning from positive and unlabelled examples. Results demonstrate that the proposed method gives good f-measure even in the presence of only small number of positive examples to learn a classifier.