摘要
现有高效识别异常引用的算法存在无法充分利用已知的标签信息或伪标签导致训练过程恶化等问题。为此提出一种融合深度强化学习和图神经网络技术的异常检测方法RACD。异常检测智能体可有效提取作者节点的异常引用特征;异常感知环境建模驱动智能体充分学习已标注数据中的异常特点,发现未标注数据中的潜在异常。通过智能体与环境的不断交互,获得最优的引用异常检测策略。在真实数据集上进行实验,其结果表明,该方法可有效检测异常学术引用。
Existing algorithms for efficiently identifying abnormal citations have problems such as not being able to make full use of known label information or deterioration of the training process caused by pseudo-labels.To deal with these issues,an anomaly detection method RACD was proposed,which integrated deep reinforcement learning and graph neural network.The anomaly detection agent could effectively extract the abnormal citation features of author nodes.The anomaly-aware environment mode-ling could drive the agent to fully learn the anomaly characteristics in the labeled data and discover potential anomalies in the unlabeled data.Through the constant interaction between agent and environment,the optimal anomaly citations detection strategy could be obtained.Results of experiments on real data sets show that RACD can effectively detect abnormal academic citations.
作者
王晓菲
朱焱
WANG Xiao-fei;ZHU Yan(School of Computing and Artificial Intelligence,Southwest Jiaotong University,Chengdu 611756,China)
出处
《计算机工程与设计》
北大核心
2024年第7期2166-2172,共7页
Computer Engineering and Design
基金
四川省科技计划基金项目(2019YFSY0032)。
关键词
图异常检测
异常学术引用
深度强化学习
图神经网络
图注意力网络
图嵌入
学术社交网络
graph anomaly detection
abnormal academic citations
deep reinforcement learning
graph neural network
graph attention network
graph embedding
academic social network
