Research And Implementation Of A Spammer Detection Method For Social Relation Attack

The rapid development of network technology promotes the mass production of information and data,and even causes information overload problems,nowadays,recommendation systems have alleviated this phenomenon to some extent and they have been used in many online platforms.Besides,with the rise of social networks,social recommendation has also become a hot research topic,with the help of social relationship information,it can alleviate the sparseness and cold start problem of recommendation algorithms.However,because the mainstream recommendation algorithms rely on the user's historical behavior data,spammers take the opportunity to inject spam information or social relationships,etc.into the recommendation system,they aim to change the normal recommendation results for gaining more profit.This behavior not only changes the normal recommendation results,reduces user satisfaction,but also seriously affects the fair online environment.Therefore,how to detect these spammers,purifying the recommended environment is an important research topic.This issue has received widespread attention from researchers.Although there are some mature methods to detect spammers,as the attack model of spammer is continuously upgraded,for example,spammers utilize social relationships to improve attack utility,and the traditional detection methods for rating attacks cannot detect these social attacks.In order to explore the spammer detection methods for social attacks and improve the detection performance,this paper first proposes a hybrid attack model with multiple fusion relationships and scores.This model is used to inject spammers into the recommended system dataset to simulate spammers in real scenes.Meanwhile,this step is important for the subsequent spammer detection work.Next is the core research content of this paper,which is a spammer detection algorithm which is suitable for dealing with social attacks.This method improves the robustness of the detection algorithm by mining users' rating features and features in the social relationship to train the detector.The main contributions of this paper are summarized as follows:(1)This thesis introduces the development trend of recommendation system and the spammer attack model in detail,classifies the spammer detection method,and analyzes the shortage of the classic attack model and the bottleneck of the spammer detection model.(2)A hybrid attack method that integrates user relationships is built,including three basic user attack methods.Combining these three attack methods with the classic scoring attack,12 kinds of mixed attacks can be constructed.This attack mode has a particularly large negative impact on social recommendation.(3)Build a heterogeneous information network,and we design multiple meta-paths to mine potential features in the network,and propose a more robust spammer detection method,which can effectively handle new types of attacks with social relationships.(4)A large number of experiments are carried out on several different datasets.The experimental results demonstrate the effectiveness of our method.(5)Developing a spammer detection system,which implements the collaborative filtering recommendation algorithm,the social recommendation algorithm,the classic attacks,the hybrid attacks and our spammer detection method.