Research And Implementation Of Semi-supervised Learning-based Social Network Anomaly Detection Scheme

With the rapid development of the Internet,surfing on online social network has become part of people's life.However,all kinds of malicious behaviors of social network abnormal users like spreading fake new,stealing user information,and controlling the public opinion also pose a huge threat to the normal users of the social network,platform managers and even the social order.Therefore,the detection of abnormal users on social networks has become an important research topic in the field of network information security.At present,the detection of abnormal users on social networks is mainly faced with the problem that the amount of user data is huge,the behavior patterns of abnormal users are constantly changing,and there is no standard abnormal measurement method.In a scenario with only a small amount of labeled data,the existing supervised learning methods have great limitations in the actual use of the detection model due to the incompleteness of training data.Although the existing unsupervised learning methods avoid the dependence on a large number of standard data sets to achieve the purpose of abnormal user detection,there is still a considerable gap between the performance of accuracy and F1 scores and supervised learning,so it cannot be actually used.In order to solve the above-mentioned problems,this paper mainly studies the technical solution of semi-supervised learning abnormal user detection by combining and using unlabeled data with labeled data.The research work of this paper is mainly composed of the following two parts: The first part is feature learning.In the feature learning stage,all social network users data(including unlabeled data and labeled data)will be input into the unsupervised learning model to extract key features.The advantage is that it not only ensures the integrity of the data features learned by the model,but also greatly reduces the high computational complexity caused by the high-dimensional sparse data.The second part is label inference.In the label inference stage,only labeled data will be input to the feedforward neural network for training,so the model can be used to update the network parameters through backpropagation,and thus obtain an efficient anomaly user detector.The author also experimented with the semi-supervised learning-based social network anomaly user detection model on public data sets.In a semi-supervised learning experiment environment,the model accuracy rate reached 90.3% and the F1 score reached 0.904,in the supervised learning experiment environment,the model accuracy rate reached 94.9%,and the F1 score reached 0.949,which were superior to the traditional anomaly user detection model.Finally,the author applies the social network abnormal user detection scheme proposed in this paper to implement a simple online social network abnormal user detection system.