I have several examples of multidimensional time series that correspond to a particular pattern (or class) and form a dictionary. Now, based on these examples I want to find if a new time series matches to one (or all) of the patterns from the dictionary or not. For that, I am performing Dynamic Time Warping (DTW) per each dimension of every example in the dictinary with the new time series and then check if the distance between them is large or not. The problem that I have is that some of the dimensions are non informative and might be different even for time series that represent the same class. Are there any methods that help to understand which dimensions of time series are informative and which not and take into account only informative ones when performing comparison?
If I correctly understand, each of your data samples can be viewed as a matrix of [time points $\times$ features]. What you could possibly do is the following approach for each class separately:
- calculate the median time-series correlation for each feature across all pairs of samples.
- A feature that is informative should have a high median, because different samples from the same class should display the same time series
- Vice versa, a feature that is uninformative should have a low median correlation
- Rank the features according to median correlation and reject the ones with low correlation
- Instead of correlation, you can of course take DTW as a distance measure as well.
Alternatively, perhaps you could also do some Supervised dimensionality reduction using LDA to map the data to a lower-dimensional discriminative space.