Abstract:Anomaly detection is fundamental yet, challenging problem with practical applications in industry. The current approaches neglect the higher-order dependencies within the networks of interconnected sensors in the high-dimensional time series(multisensor data) for anomaly detection. To this end, we present a self-adapting anomaly detection framework for joint learning of (a) discrete hypergraph structure and (b) modeling the temporal trends and spatial relations among the interdependent sensors using the hierarchical encoder-decoder architecture to overcome the challenges. The hypergraph representation learning-based framework exploits the relational inductive biases in the hypergraph-structured data to learn the pointwise single-step-ahead forecasts through the self-supervised autoregressive task and predicts the anomalies based on the forecast error. Furthermore, our framework incentivizes learning the anomaly-diagnosis ontology through a differentiable approach. It derives the anomaly information propagation-based computational hypergraphs for root cause analysis and provides recommendations through an offline, optimal predictive control policy to remedy an anomaly. We conduct extensive experiments to evaluate the proposed method on the benchmark datasets for fair and rigorous comparison with the popular baselines. The proposed method outperforms the baseline models and achieves SOTA performance. We report the ablation studies to support the efficacy of the framework.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to perform anomaly detection, control, and optimization in industrial data. Specifically, current methods ignore the high - order dependencies between interconnected sensors in the network when dealing with anomaly detection in high - dimensional time series (multi - sensor data). To this end, the authors propose a recommendation system framework based on hypergraph learning, aiming to jointly learn (a) discrete hypergraph structures and (b) model time trends and spatial relationships using a hierarchical encoder - decoder architecture to overcome these challenges. ### Main contributions of the paper 1. **Hypergraph structure learning technology**: Infer dynamic hypergraphs through similarity learning methods to achieve a structured spatio - temporal representation of multi - sensor data. 2. **Novel attention - mechanism hypergraph convolution technology**: Perform spatial hypergraph filtering operations. 3. **Novel local hypergraph pooling and unpooling technologies**: Perform down - sampling and up - sampling operations on hypergraphs of arbitrary structures. 4. **Novel single - stage hierarchical attention - mechanism encoder - decoder technology**: Learn high - order representations of hypergraph - structured data. ### Framework overview The framework (HgAD) proposed in the paper contains the following modules: - **Hypergraph structure learning (HgSL) module**: Learn the implicit high - order structural patterns in multi - sensor data and capture the internal dependencies of interconnected sensors in the network. - **Encoder - decoder (HgED) module**: Operate on the hypergraph topology, learn hierarchical representations, and simultaneously encode the time trends and spatial relationships of multiple IoT sensors, preserving the global topological properties of the hypergraph. - **Hypergraph prediction (HgF) module**: Utilize hyper - node representations to predict the one - step forward prediction values of sensors. - **Hypergraph deviation (HgD) module**: Mark possible abnormal events by comparing expected and actual trends, use the prediction error as the criterion for anomaly detection, and provide supervision information for the HgSL module. ### Anomaly detection process 1. **Hypergraph representation**: Convert multivariate continuous time - series data into fixed - length inputs and generate sliding - window data. 2. **Hypergraph structure learning**: Dynamically optimize the sensor topology by learning sensor embeddings and implicitly learn task - related structural relationships. 3. **Hyper - node position encoding**: Encode the position information of hyper - nodes to make the model position - aware. 4. **Hypergraph encoder - decoder**: Learn high - order dependencies and hypergraph structural features through hypergraph convolution, local pooling, and unpooling operations. 5. **Hypergraph prediction**: Predict multi - sensor values and minimize the mean - squared error between the predicted and actual values. 6. **Hypergraph deviation**: Calculate robustly normalized anomaly scores to accurately locate abnormal points in multi - sensor data. ### Experimental results The paper verifies the effectiveness of the proposed method through experiments. The main research questions include: - **Accuracy (RQ1)**: How does the proposed method perform on the anomaly detection task? - **Ablation study (RQ2)**: How much do the individual modules of the proposed method contribute to the improvement of the model performance? - **Interpretability (RQ3)**: How does the proposed method provide anomaly diagnosis and suggest measures to avoid anomalies? ### Benchmark datasets The paper is evaluated on a variety of public datasets, including datasets from real - world water treatment facilities such as SWaT and WADI. The experimental results show that the proposed method outperforms the baseline methods on the benchmark datasets and demonstrates the effectiveness of each module in the ablation study. Through these contributions, the method proposed in the paper not only achieves a significant performance improvement in anomaly detection but also provides in - depth analysis and effective solutions for abnormal events.

Hypergraph Learning based Recommender System for Anomaly Detection, Control and Optimization

Accurate Anomaly Detection Leveraging Knowledge-enhanced GAT

RustGraph: Robust Anomaly Detection in Dynamic Graphs by Jointly Learning Structural-Temporal Dependency

Graph Neural Network-Based Anomaly Detection in Multivariate Time Series

Multivariate Time Series Anomaly Detection via Dynamic Graph Forecasting

Hyperbolic Self-supervised Contrastive Learning Based Network Anomaly Detection

Multivariate Time-Series Anomaly Detection via Graph Attention Network

AD-Graph: Weakly Supervised Anomaly Detection Graph Neural Network

Extreme Value Modelling of Feature Residuals for Anomaly Detection in Dynamic Graphs

Learning Graph Neural Networks for Multivariate Time Series Anomaly Detection

A Hierarchical Approach to Conditional Random Fields for System Anomaly Detection

Self-Supervised Variational Graph Autoencoder for System-Level Anomaly Detection

Heterogeneous Hypergraph Embedding for Recommendation Systems

An Evaluation of Anomaly Detection and Diagnosis in Multivariate Time Series

Automatic Hypergraph Generation for Enhancing Recommendation with Sparse Optimization

Hybrid Anomaly Detection via Multihead Dynamic Graph Attention Networks for Multivariate Time Series

Low-rank on Graphs plus Temporally Smooth Sparse Decomposition for Anomaly Detection in Spatiotemporal Data

Asymptotic Consistent Graph Structure Learning for Multivariate Time-Series Anomaly Detection

An Unsupervised Anomaly Detection Approach Using Energy-Based Spatiotemporal Graphical Modeling

Context-Dependent Anomaly Detection with Knowledge Graph Embedding Models

Application of Improved Asynchronous Advantage Actor Critic Reinforcement Learning Model on Anomaly Detection