Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Graph-Augmented LSTM for Forecasting Sparse Anomalies in Graph-Structured Time Series

Sneh Pillai

TL;DR

The paper tackles anomaly detection in graph-structured, multivariate time series with sparse anomalies by introducing Graph-Augmented LSTM, which propagates neighbor information during forecasting. Anomalies are detected from forecast residuals via per-node thresholds, with explicit modeling of the graph topology to capture cross-series dependencies. Empirical results on Yahoo S5 and METR-LA show state-of-the-art F1-scores, with significant gains over LSTM-only, ARIMA, and Prophet, and ablations confirming the critical role of accurate graph structure. The work demonstrates the practical value of graph-informed forecasting for reliable anomaly detection in complex systems and outlines avenues for dynamic graph learning and more expressive graph neural architectures.

Abstract

Detecting anomalies in time series data is a critical task across many domains. The challenge intensifies when anomalies are sparse and the data are multivariate with relational dependencies across sensors or nodes. Traditional univariate anomaly detectors struggle to capture such cross-node dependencies, particularly in sparse anomaly settings. To address this, we propose a graph-augmented time series forecasting approach that explicitly integrates the graph of relationships among time series into an LSTM forecasting model. This enables the model to detect rare anomalies that might otherwise go unnoticed in purely univariate approaches. We evaluate the approach on two benchmark datasets - the Yahoo Webscope S5 anomaly dataset and the METR-LA traffic sensor network - and compare the performance of the Graph-Augmented LSTM against LSTM-only, ARIMA, and Prophet baselines. Results demonstrate that the graph-augmented model achieves significantly higher precision and recall, improving F1-score by up to 10% over the best baseline

Graph-Augmented LSTM for Forecasting Sparse Anomalies in Graph-Structured Time Series

TL;DR

The paper tackles anomaly detection in graph-structured, multivariate time series with sparse anomalies by introducing Graph-Augmented LSTM, which propagates neighbor information during forecasting. Anomalies are detected from forecast residuals via per-node thresholds, with explicit modeling of the graph topology to capture cross-series dependencies. Empirical results on Yahoo S5 and METR-LA show state-of-the-art F1-scores, with significant gains over LSTM-only, ARIMA, and Prophet, and ablations confirming the critical role of accurate graph structure. The work demonstrates the practical value of graph-informed forecasting for reliable anomaly detection in complex systems and outlines avenues for dynamic graph learning and more expressive graph neural architectures.

Abstract

Detecting anomalies in time series data is a critical task across many domains. The challenge intensifies when anomalies are sparse and the data are multivariate with relational dependencies across sensors or nodes. Traditional univariate anomaly detectors struggle to capture such cross-node dependencies, particularly in sparse anomaly settings. To address this, we propose a graph-augmented time series forecasting approach that explicitly integrates the graph of relationships among time series into an LSTM forecasting model. This enables the model to detect rare anomalies that might otherwise go unnoticed in purely univariate approaches. We evaluate the approach on two benchmark datasets - the Yahoo Webscope S5 anomaly dataset and the METR-LA traffic sensor network - and compare the performance of the Graph-Augmented LSTM against LSTM-only, ARIMA, and Prophet baselines. Results demonstrate that the graph-augmented model achieves significantly higher precision and recall, improving F1-score by up to 10% over the best baseline

Paper Structure

This paper contains 18 sections, 2 equations, 6 figures, 1 table.

Table of Contents

  1. Introduction
  2. Contributions and Methodology
  3. Methodology
  4. Datasets
  5. Models for Anomaly Detection
  6. Graph-Augmented LSTM (Ours).
  7. LSTM-Only Baseline.
  8. ARIMA.
  9. Prophet.
  10. Evaluation Metrics
  11. Graph Ablation Experiment
  12. Results
  13. Overall Detection Performance
  14. Detection Example.
  15. Node-Level Analysis
  16. ...and 3 more sections

Figures (6)

  • Figure 1: Per-node anomaly counts in train and test data, illustrating severe class imbalance.
  • Figure 2: Threshold tuning results for each node: best F1-score achieved and corresponding optimal threshold.
  • Figure 3: Forecast vs actual for a Yahoo S5 time series, with detected anomalies (Graph-LSTM) and ground truth anomalies.
  • Figure 4: Per-Node F1 Improvement vs Node Degree on METR-LA. Degree indicates the number of neighboring sensors in the road network.
  • Figure 5: Node-wise F1 comparison between real graph and random graph on METR-LA.
  • ...and 1 more figures