Unsupervised Scalable Representation Learning for Multivariate Time Series

TL;DR

This paper introduces an unsupervised, scalable method for learning universal embeddings of multivariate time series, addressing variability in length and sparse labels, and demonstrating strong transferability and practicality.

Contribution

It proposes a novel scalable unsupervised approach combining causal dilated convolutions with triplet loss for effective time series representation learning.

Findings

High-quality embeddings demonstrated through extensive experiments

Effective transferability of learned representations across tasks

Scalability to variable-length and multivariate time series

Abstract

Time series constitute a challenging data type for machine learning algorithms, due to their highly variable lengths and sparse labeling in practice. In this paper, we tackle this challenge by proposing an unsupervised method to learn universal embeddings of time series. Unlike previous works, it is scalable with respect to their length and we demonstrate the quality, transferability and practicability of the learned representations with thorough experiments and comparisons. To this end, we combine an encoder based on causal dilated convolutions with a novel triplet loss employing time-based negative sampling, obtaining general-purpose representations for variable length and multivariate time series.