NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:2607
Title:Unsupervised Scalable Representation Learning for Multivariate Time Series


		
This paper presents a new technique for time series embedding for multivariate time series of differing lengths. The method encodes the time series with a stack of dilated causal convolutions and uses "triplet-loss" function that has been adapted to the time series domain. Overall, the reviewers found the combination of these pieces novel for the problem the authors were trying to solve. Additionally, the experiments were quite extensive and demonstrated as well as could be asked for that the method performs well and is useful. There were a few criticisms of the evaluation such as comparing to other tasks and using KNN instead of SVMs for classification. The authors seem to have addressed these issues to the reviewers' satisfaction. The authors should incorporate the useful feedback that the reviewers provided in the camera-read version.