Graph Neural Network-Based Anomaly Detection in Multivariate Time Series
In this presentation, a novel approach to anomaly detection using Graph Neural Networks (GNNs) is introduced. This method models multivariate time series as a graph, learns sensor relationships and detects anomalies by identifying deviations from these learned patterns. By capturing both temporal dependencies and sensor interactions, proposed approach enhances anomaly detection, offering superior performance compared to traditional techniques.
GraphAny
GraphAny is a groundbreaking foundation model for node classification on any graph, addressing the limitations of existing models that struggle to generalize across different graph structures and feature spaces. This innovative approach leverages a novel architecture consisting of LinearGNNs and an attention mechanism that learns to combine their predictions, enabling effective inference on new graphs without the need for retraining. By learning attention scores for each node based on entropy-normalized distance features, GraphAny demonstrates remarkable generalization capabilities, surpassing traditional methods in various node classification tasks.