I can't get any further with my segmentation/clustering/classification problem and need help in choosing the right tools, or rather in leading me to the right problem definition.
I have a single long time series in which I would like to find clusters that break it down into different crisp segments (chopping). There are no initial labels, so it should be an unsupervised learning task. Later it should be possible to classify the cluster/segments on the basis of a small selection of meaningful features.
You may see that I am not that experienced in the subject of data mining and caught between very different tasks: segmentation, clustering, feature selection and classification.
My approach:
- Extracting features from an overlapping sliding window for a certain time step (~ 300 features, from distributions over FFT and WLT decomposition, entropy etc.)
- Looking for features that describe 'stable' states, i.e that are well suited for clustering in the feature space but in respect of time.
- Based on the selected features a classifier - operating in a slinding window again - should recognize the defined clusters by the seleced features.
Especially with point 2) I don't get any further. How do I rank meaningful features for clustering subspaces/sections? Meaningful here means some kind of stability over time. My idea was to use a changing point detection and to prioritize the features with the fewest recognized changing points; or to use other segmentation algorithms (buttom-up) to look where the fewest segments occur with a defined error. Or are there suitable feature selection algorithms that adress this problem?
The plot shows a first ranking according to the laplacian score. It brings the very good feature Op: 293 to the top. Are there good reasons for using this method? I have chose it more by chance among the unsupervised FSS methods
In general, also the question of whether my approach makes sense at all, if I mix up segmentation and clustering? I really need thoughts / exchange at the moment.
