Research On Clustering And Anomaly Detection Based On Deep Learning

Clustering and anomaly detection,which are classic data analysis algorithms,are widely used in the data mining field.Anomaly detection can be regarded as a preprocessing method of clustering,and clustering can be regarded as an intermediate step of anomaly detection,the two complement each other.However,traditional clustering and anomaly detection algorithms rely on manual extraction of data features.When algorithms deal with complex and diverse data,they require a lot of manpower for data preprocessing.With the development of deep learning in the feature engineering field,clustering and anomaly detection algorithms based on deep learning have received extensive attention.In this dissertation,we focus on the problems of clustering and anomaly detection algorithms based on deep learning,and provide effective data analysis methods for the data mining and management.The main contributions are as follows:(1)Deep graph clustering.Firstly,to overcome the drawbacks that the existing algorithms ignore node attributes when extracting graph features,and fail to capture the clustering structure information,this dissertation proposes a deep graph clustering algorithm based on graph neural networks.The algorithm constructs a graph neural network to simultaneously consider the node attributes and topology in the graph,and finds the similar relationship between the graphs according to the graph kernel and pseudo-label information,which improves the effectiveness of graph clustering.Secondly,to overcome the drawback that graph neural networks extracting scalar graph features cannot fully consider graph attribute information,this dissertation designs a graph clustering capsule neural network to capture vectorized graph features.The capsule network parameters and clustering assignment parameters are updated through an iterative optimization strategy,which enhances the representation ability of the network.(2)Deep multimodal clustering.The existing deep multimodal clustering algorithms first learn data features and then cluster the data features.The learned data features are not suitable for clustering tasks.Moreover,the network structure of existing algorithms cannot satisfy the constraints of clustering.To address these problems,this dissertation proposes a deep multimodal clustering algorithm based on end-to-end pattern.To make the model directly output the clustering results,this algorithm adjusts the decoder structure and minimizes the distance between the reconstructed data and the original data.To fully learn the consistency information between modalities,this algorithm introduces a multimodal feature fusion module to narrow the gap between the fusion feature and different modal features,which effectively improves the clustering results.(3)Deep anomaly detection.Firstly,to overcome the drawbacks that the inference-based algorithms adopting the adversarial training constraint objective function cannot capture the clustering structure of normal data,and ignore the marginal distribution of normal data,this dissertation proposes an anomaly detection algorithm based on auxiliary task and adversarial training.The clustering structure of normal data is learned by classifying normal data and auxiliary data,and a deep adversarial training model is constructed to capture the marginal distribution of normal data.Secondly,to overcome the drawbacks that algorithms based on oneclass classifier using one-class constraint objective function lose text content information,and scalar attention weights cannot focus on word dimensional attributes,this dissertation proposes a textual anomaly detection algorithm based on attention mechanism.This algorithm designs the attention mechanism and adversarial training method to capture the textual semantic information,and constructs dimensional transformation matrices to extract the dimensional attributes of polysemous words,which improves the effectiveness of textual anomaly detection.