Vous êtes ici

Shapelet-neural-networks for weakly supervised problems

Equipe et encadrants
Département / Equipe: 
Site Web Equipe: 
Directeur de thèse
Guillaume Gravier
Co-directeur(s), co-encadrant(s)
Malinowski Simon
NomAdresse e-mailTéléphone
Malinowski Simon
Sujet de thèse

In the time series analysis domain, very efficient methods have been developed recently for supervised tasks (e.g. classification).
Amongst them, shapelet-based models are known to be efficient both in terms of accuracy and complexity [1,2,3]

However, in a wide range of applications, very little amount of supervised information is available which prevents
from using the above-cited methods directly.

Recently, some efforts have been dedicated to design unsupervised methods for time series analysis [4,5,6,7]
These works mainly focus on the particular task of time series clustering.

The aim of this thesis is to explore several weakly supervised tasks for time series analysis.
For that purpose, we will be particularly interested in bridging the gap between shapelets and neural networks,
in order to learn efficient representations for time series in a weakly supervised context.

In [6], we designed LDPS, a model combining shapelet and siamese networks in order to embed time series in
a space where Euclidean distance mimics a widely used similarity measure for time series analysis (DTW).
We aim at extending this framework to the following tasks:

 - time series indexing under DTW: this task is known to be very challenging [8]. We expect that an anytime extension
 of the LDPS framework would be of great help for this task.

 - metric learning and semi-supervised learning : we will be interested in extending the LDPS framework for situations where only a few labels
  are available (semi-supervised task) or where supervised information is available as must-link/cannot-link constraints (as in the metric learning framework)

 Other tasks (eg. domain adaptation) could also be considered.

The methods developed in this thesis will be applied to multimedia data in the context of efficient document retrieval.


[1] Hills, J., Lines, J., Baranauskas, E., Mapp, J., Bagnall, A.: Classification of
time series by shapelet transformation. DMKD 28(4) (2014)

[2] Grabocka, J., Schilling, N., Wistuba, M., Schmidt-Thieme, L.: Learning
time-series shapelets. In: Proc. KDD (2014)

[3] Ye, L., Keogh, E.: Time series shapelets: a new primitive for data mining.
In: Proc. KDD (2009)

[4] Zakaria, J., Mueen, A., Keogh, E.: Clustering time series using
unsupervised-shapelets. In: Proc. ICDM (2012)

[5] Zhang, Q., Wu, J., Yang, H., Tian, Y., Zhang, C.: Unsupervised feature
learning from time series. In: Proc. IJCAI (2016)

[6] Lods, A., Malinowski S., Tavenard R.,  Learning DTW-preserving shapelets. In Proc. IDA (2017)

[7] Cuturi, M., Blondel M.:Soft-DTW: a differentiable loss function for time series. In Proc. ICML (2017)

[8] Keogh, E., Ratanamahatana, C.A.: Exact Indexing of Dynamic Time Warping.
KAIS 7 (2005)

Début des travaux: 
Mots clés: 
Machine Learning, Time series, Representation learning
IRISA - Campus universitaire de Beaulieu, Rennes