Group Attack Detection Algorithm In Recommender Systems Based On Graph Neural Network

In order to create greater profits in a short period of time and avoid being discovered by attack detection methods,malicious users start to commit fraud.Group attack is more harmful than traditional single-user attack.However,the existing detection methods mostly focus on detecting single attacking users,and there are problems in ignoring the hidden characteristics of users and the generalization ability of the detection model is not strong.In order to overcome these limitations,this paper proposes a group attack detection algorithm for recommendation systems based on graph neural networks.Firstly,the existing algorithms ignore the hidden features of users and the imbalance of positive and negative samples leads to poor detection results.A group attack detection algorithm for recommendation system based on graph convolutional neural network is proposed.The algorithm uses historical user comment data to construct a heterogeneous graph containing user,product,and comment data.And get the node sequence of related users from multi-angle projection.On this basis,Glo Ve is used to learn user hidden features,generate user feature vectors,and use clustering algorithm to obtain candidate groups.Then,the graph convolutional neural network model is used to detect the attack group based on the user relationship network and user feature vector of the candidate group.Secondly,in order to solve the problem that the previous detection methods used a large amount of labeled data as the basis for detection and the low difference between the candidate groups.A semi-supervised graph attention network recommendation system group attack detection algorithm is proposed.This algorithm uses the candidate groups obtained by the above clustering.Then the user's local structural features are integrated into the model.The semi-supervised graph attention network model is used to aggregate user characteristics and user structure characteristics based on weight bias.After multiple iterations,the group vector of the candidate group is obtained and detected.In the detection model training process,a dual-model-guided training method is designed to make full use of labeled and unlabeled data to obtain the final attack group.Finally,for the methods mentioned in this article.Experiment on Amazon and Netflix datasets.The comparison with existing algorithms shows that the algorithm proposed in this paper can effectively detect group attacks.