Anomaly Detection In Attributed Networks Based On Deep Autoencoder

Attributed networks have a wide range of applications in many fields,such as social media analysis,criminal behavior detection,biomedical diagnosis,and so on,which has attracted increasing research interests in recent years.One of the research hotspots in attribute network analysis is anomaly detection.With the increasing scale,complexity,and sparseness of real-world networks,traditional methods based on graph analysis become difficult to meet the requirements of large-scale data processing.Recently,applying Network Representation Learning(NRL)to graph anomaly detection has attracted the attention of many researchers.The advantage of this kind of method is that it can perform both key feature extraction and anomaly detection simultaneously.The result of NRL is a low-dimensional embedding of the target network,which extracts the significant features and helps simplify the subsequent computing processes.Also,the learned embedding results can preserve the original distance between nodes in the lowdimensional space.So we can effectively discover the anomalies with the reconstruction error from low-dimensional to high-dimensional space.Despite the current success in this area,there are still some challenges remain.For instance,the particularity of the anomalies may affect the quality of the embedding of normal nodes,causing the inaccuracy of anomaly detection results.Besides,the non-linear interactions between nodes and the high-dimensionality of attribute networks also need to be addressed by the model.Existing methods cannot achieve a balance between efficiency and availability of results.Thus,this thesis proposes an unsupervised anomaly detection model named AEAD(Auto Encoder based Anomaly Detection).The main content of this thesis consists of the following points:1)This thesis introduces the theoretical basis and related techniques of anomaly detection and NRL,and emphatically analyzed several methods that combined these two concepts.2)The AEAD model proposed in this thesis base on multi-input deep autoencoders,which process the structure and content information of the network concurrently.Taking the advantages of depth,the model can deal with the high-dimensionality and capture the underlying non-linear characteristics of the network.The AEAD model detects anomalies depending on the reconstruction errors generated by the autoencoders while minimizing the impact of anomalies.More specifically,based on existing research works,an improved loss function is designed for AEAD,which introduces anomaly scores as penalty factors for potential anomalies.As a consequence,the model tends to learn the features of normal nodes,resulting in a higher accuracy of anomaly detection and a more robust low-dimensional representation of the network.To demonstrate the effectiveness of the proposed AEAD model,this thesis conducted comparative experiments on both anomaly detection and downstream mining tasks on three different real-world network datasets.The experimental results show that the AEAD model has improved performance on both anomaly detection and network representation learning tasks.