What are the primary options regarding classification problems with time series with more than one feature and different lengths? So far I've read of k-means with dtw, but haven't seen it applied to more than one feature.
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
Dynamic Time Warping (DTW) could also be generalized to the case of multivariate time series, as in your case. There are at least two ways of generalization:
- DTW$_I$ where each feature/dimension is treated independently (the subscript $I$);
- DTW$_D$ where each feature/dimension is treated dependently (the subscript $D$).
I would not call $k$-means a classification algorithm: it is a clustering one.
The paper that presents the two can be found here.
Finally, as for classification of multivariate time series, you could use the similarity score provided by either DTW, and use a similarity/distance/instanced-based algorithm for clustering such as $k$-NN (e.g., $1$-NN). If you are interested in the state-of-the-art for classification of multivariate time series, you should read this paper.
