Harnessing Feature Clustering For Enhanced Anomaly Detection With Variational Autoencoder And Dynamic Threshold

TL;DR

The paper tackles anomaly detection in Arctic climate time series by introducing a Cluster-VAE that clusters features via Pearson correlation, processes clusters with LSTM-VAE encoders, and applies dynamic thresholding via POT to detect localized anomalies. It also incorporates feature perturbation to identify drivers of extreme snowmelt and emphasizes region-specific explainability. The approach outperforms several baselines on benchmark datasets and aligns with NSIDC ground truth, demonstrating both improved detection and interpretability. This framework enhances monitoring of climate extremes and supports policy-relevant insights for Arctic regions.

Abstract

We introduce an anomaly detection method for multivariate time series data with the aim of identifying critical periods and features influencing extreme climate events like snowmelt in the Arctic. This method leverages the Variational Autoencoder (VAE) integrated with dynamic thresholding and correlation-based feature clustering. This framework enhances the VAE's ability to identify localized dependencies and learn the temporal relationships in climate data, thereby improving the detection of anomalies as demonstrated by its higher F1-score on benchmark datasets. The study's main contributions include the development of a robust anomaly detection method, improving feature representation within VAEs through clustering, and creating a dynamic threshold algorithm for localized anomaly detection. This method offers explainability of climate anomalies across different regions.

Figures (4)

• Figure 1: Model Framework
• Figure 2: Greenland map with area studied shaded in blue
• Figure 3: The feature importance ranking computed using feature perturbation. A) 2019, B) 2021
• Figure 4: The year to year trend of observed anomalies from 1941 to 2022

Theorems & Definitions (2)

• Definition 1
• Definition 2