ニューラルネットワークによる時系列モチーフ発見

抄録

Matrix profile is widely recognized for its effectiveness in tasks such as similarity search, motif discovery, and anomaly detection within time series data. However, its practical application is hindered by certain limitations, including its exhaustive nature and the need for interpretation. To address these issues, we propose an autoencoder architecture for time series data that outputs a subsequence similarity map. The proposed method provides an adaptive neural network solution for motif and discord discovery.