Entropy-based dynamic graph embedding for anomaly detection on multiple climate time series

Gen Li,Jason J Jung

doi:10.1038/s41598-021-92973-8

Abstract

Abnormal climate event is that some meteorological conditions are extreme in a certain time interval. The existing methods for detecting abnormal climate events utilize supervised learning models to learn the abnormal patterns, but they cannot detect the untrained patterns. To overcome this problem, we construct a dynamic graph by discovering the correlation among the climate time series and propose a novel dynamic graph embedding model based on graph entropy called EDynGE to discriminate anomalies. The graph entropy measurement quantifies the information of the graphs and constructs the embedding space. We conducted experiments on synthetic datasets and real-world meteorological datasets. The results showed that EdynGE model achieved a better F1-score than the baselines by 43.2%, and the number of days of abnormal climate events has increased by 304.5 days in the past 30 years.

Highlights

Abnormal climate event is that some meteorological conditions are extreme in a certain time interval
We discriminate the anomaly by using the entropy-based dynamic graph embedding model (EDynGE) model
local outlier factor (LOF) and isolation forest (IF) methods can control the ratio of anomalies detected so that the performance of these two methods are not affected by selecting the different ratio of anomalies

Summary

Introduction

Abnormal climate event is that some meteorological conditions are extreme in a certain time interval. The existing methods for detecting abnormal climate events utilize supervised learning models to learn the abnormal patterns, but they cannot detect the untrained patterns To overcome this problem, we construct a dynamic graph by discovering the correlation among the climate time series and propose a novel dynamic graph embedding model based on graph entropy called EDynGE to discriminate anomalies. The semi-supervised learning methods utilize a few labeled data to fit the models and detect the abnormal climate events on the unlabeled data These methods are commonly constructed by using the autoencoder. Climate events comprise the multiple meteorological data, and the correlation between these data plays an essential role in anomaly detection To address this problem, we propose using a graph to model the correlation among multiple time series. It is formulated as|e(Gi) − e(Gi−1)| > θ or|e(Gi) − e(Gi+1)| > θ , where e(Gi) is the entropy of the climate event Gi and θ is the threshold for detecting the abnormal climate event

Objectives

Methods

Results

Conclusion