Unsupervised Discovery in Multivariate Time Series

A new study introduces a method for discovering relationships in multivariate time series, based on latent structural similarity networks. The approach aims to construct a relational hypothesis graph over entities without assuming linearity or stationarity.

Method Architecture

The proposed method uses an unsupervised sequence-to-sequence autoencoder to learn window-level representations. These representations are then aggregated into entity-level embeddings. A sparse similarity network is induced by thresholding a similarity measure in the latent space.

Objectives and Applications

The resulting network serves as an analyzable abstraction, compressing the pairwise search space and exposing candidate relationships for further investigation. The goal is not optimization for prediction or trading, but the analysis of underlying relationships. The framework was tested on a cryptocurrency dataset, demonstrating the ability to induce a coherent network structure.