An ensemble approach to detect review spam using hybrid machine learning technique

Online reviews are becoming one of the vital components of e-commerce in recent years as so many people consider having different opinions prior to buying online products or apprehending any online service. Nowadays, in the era of web 2.0, it is completely understandable that people rely on online reviews more than ever while taking a decision. However, guaranteeing the authenticity of these sensitive and valuable information is hardly visible. Due to fulfill some immoral benefits, many people post fake review or fabricated opinion to uphold or devalue a certain product or service which certainly hampers the ingenuousness of the real fact. To detect fake reviews, many methodologies were introduced by harvesting the obvious content features, rating consistency, empirical conditions, helpfulness voting etc. The most of them are supervised models which mostly rely on pseudo fake reviews and the scarcity of good quality large-scale labeled dataset is still a hindrance. In this paper, we introduce an ensemble learning approach which combines two different types of learning methods (active and supervised) by creating a hybrid dataset of both real-life and pseudo reviews. This model holds 3 different filtering phases that is based on KL and JS distance, TF-IDF features and n-gram features of the review content. It achieves phenomenal results while working on almost 3600 reviews from different domains. In the best case, the precision, recall and f-score are above 95% and the accuracy it achieved is slightly above 88%. In the process, about 2000 reviews were manually labeled. After evaluating and comparing the results with other successful methods, it is quite clear that this detecting method is efficient and very promising.