Research On Spammer Detection In Online Shopping Websites

With its rising popularity, online shopping is one of the hottest Internet-based web services. The valueable user information existed in the online shopping websites makes the spammers excited in writing untruful comments to improve or deprove the target product.Spammers try their best to influence the rate of target products and make normal users buy the products with low quality. In order to avoid detecting, spammers manipulate a large numbers of spot accounts to act like a normal buyer, making detection face more challenge.Therefore, building intelligent and effective model to identify e-shopping spammers is very urgent and important. The main work of this thesis includes the following two parts:First, this thesis focuses on the evaluation of effective behavior features of e-shopping websites spammers. Until now, the current research work shared some commen characters.Among all the features employed in detection, behavioral features which are represented by the behaviors of spammer can not be easily covered.The efficience of behavioral features outstand other kinds of methods in order to effectively spot the spammers. This part of thesis summarizes and analyzes different kinds of behavior features and models to conducts all the effective features on two typical real-world review datasets: Taobao.com and Amazon.com. Experiment results prove that different features showed different ability to identify the hidden e-shopping spammers.Second, this thesis pays more attention on find the large spammer groups based on their relation graph built by the common targets. This part proposes a novel angle to address these kinds of large spammer groups as anomaly sub-graphs of the user relation graph of e-shopping websites users. An algorithm is proposed to detect the anomaly sub-graphs representing the large spammer groups by analyzing its spectrum features.Experiments on two popular real-life review datasets demonstrate the effectiveness of the proposed method which outperforms the existing methods to identify large spammer groups. The model proposed in this thesis provides handful assistance to manage e-shopping website, analyze users’ shopping habit and address the advertisement events.