A self-supervised graph embedding approach based on hierarchical projection network called MeghenNet(multi-view heterogeneous graph projection network) was introduced to learn low-dimensional representations from multiple views. The concept of multiple-view heterogeneous graphs was defined to explicitly allow the model to simultaneously collect information from multiple data sources for modeling heterogeneous graphs. A hierarchical attention projection that involves a cross-relation projection to extract semantics information within each view was employed, followed by a cross-view projection to aggregate contextual information from other views. The mutual information loss function between each view embedding and the global embedding was computed to ensure the information consistency across views. Experimental results on several real-world datasets demonstrate that the proposed method outperforms state-of-the-art approaches when handling multi-view heterogeneous graphs.

Addressing the issue of anomaly detection on missing multivariate time series data in real IoT (Internet of things) environments, a novel method on multivariate time series anomaly detection algorithm intergrated with graph embedding of missing information was proposed. Using a joint learning framework of pre-interpolation and anomaly detection task fusion, a GNN (graph neural network) pre-interpolation module based on time series Gaussian kernel function was designed to realize the joint optimization of pre-interpolation and anomaly detection task. A graph structure learning method for embedding missing information in time series data was proposed, using graph attention mechanism to fuse missing information masking matrix and spatiotemporal feature vectors, effectively modeling the potential connections of missing data distribution in multivariate time series. The performance of the algorithm was verified on real IoT sensor datasets. Experimental results prove that the proposed method significantly outperform the mainstream two-stage methods on the task of missing multivariate time series anomaly detection. The comparative experiment of the pre-interpolation module fully prove the effectiveness of the GNN pre-interpolation layer based on the Gaussian kernel function.