Metric Learning for Temporal Sequence Alignment

Damien Garreau (1, 2) , Rémi Lajugie (1, 2) , Sylvain Arlot (1, 2) , Francis Bach (1, 2)

1 SIERRA - Statistical Machine Learning and Parsimony
2 DI-ENS - Département d'informatique - ENS Paris

In this paper, we propose to learn a Mahalanobis distance to perform alignment of multivariate time series. The learning examples for this task are time series for which the true alignment is known. We cast the alignment problem as a structured prediction task, and propose realistic losses between alignments for which the optimization is tractable. We provide experiments on real data in the audio to audio context, where we show that the learning of a similarity measure leads to improvements in the performance of the alignment task. We also propose to use this metric learning framework to perform feature selection and, from basic audio features, build a combination of these with better performance for the alignment.

Keywords: machine learning; structured prediction; dynamic time warping; Frank-Wolfe; optimization; score following; music information

Domains: Machine Learning [cs.LG]

Damien Garreau, Rémi Lajugie, Sylvain Arlot, Francis Bach. Metric Learning for Temporal Sequence Alignment. Advances in Neural Information Processing Systems 27 (NIPS 2014), Dec 2014, Montréal, Canada. ⟨hal-01062130⟩