Shooting review spam with a weakly supervised approach and a sentiment-distribution-oriented method

Untruthful opinions, ballot-stuffing or bad-mouthing online commodities, are challenging to identify because of two prominent obstacles, i.e., lack of ground-truth annotations and cracking deceptive sentiment along review contexts. To rise to these challenges, inspired by a recent algorithm called Learning with Label Noise, we first recruit volunteers to write annotated reviews and then label more unannotated public reviews with a neighborhood graph. Furthermore, based on statistical analysis, we introduce a Sentiment-Distribution-Oriented Clustering (SDOC) method to ferret review spam out, in which product usage aspects and their sentiment polarities are highlighted. Evaluations and comparisons with several state-of-the-art approaches indicate that SDOC is effective and outperforms them with statistical significance. We have also arrived at an interesting conclusion, i.e., genuine reviewers’ feelings tend to fluctuate across different product aspects, whereas spammers always have uniform sentiments along aspects.