I'm Working on clustering task on Event-related potentials(ERP) data, I used many Clustering algorithms such as (K-means, Hierarchical, SOM, FCM and etc.)to do that. The problem is, except using "correlation" based similarity measurement the other measurements find different objects (different from the view of neurosciences) in same cluster or similar objects in different clusters, How to deal which this problem, any experience or suggestion?
(Enclosed sample is by using SOM for clustering)
Any time series contains two key parts : Oscillation and rhythm.
Oscillations are characterized by their amplitude and phase. Rhythm are characterized by frequency. Most methods are used to process the time data according to these directions.
Correlation is complex concept, you can use simple correlation between two time series to compare their shape similarity or short delay in their time to calculate the correlated.
We often calculate the coherence in their phase and wavelet-based methods to study dynamical interrelations between brains signals.
When we talk to frequency, we often use coherence, bicoherence and mult-channel bicoherence to get their similarity matrix.
In conclusion, too many methods and ways to get so-called correlation matrix, but Each method corresponds to the different meanings. The arguments above are personal perspectives , welcome comments and correction.
Hope this a little help for you
Dear Reza,
you want to cluster ERPs? Then, what are the individual exemplars in your distribution. Are these different subjects in one and the same condition? Or parametrically varied conditions? Many different stimuli ... It pretty much depends in which time and frequency domain you expect the relevant process. Then correlation or coherence might well work. Can you supply the details?
Have you thought about representational similarity analysis? This would make sense if e.g. different visual stimuli are your exemplars. Here you do not cluster but try to understand which aspects of stimulus space are mapped on response space.
A more standard approach of ERP analysis is checking for statistical differences. Here threshold free cluster analysis is often used. But this is not what you mean?
Best, Peter
Dear Tianyi
Thank you so much!
Actually, I have used correlation based similarity for ERPs clustering task, but I need more methods. The research motivation is to find time samples with similar topography, So I need to know for given ERP dataset which time samples are similar based on topography. In other words, they may share similar brain activity in practice.
Could you please tell me do the " bicoherence and multi-channel bicoherence " outcome is cluster or similar frequency bins, what about topography?
Sincerely,
Dear Peter,
Thanks a lot!
You are right, I am working on ERPs, it could be different types of experiments. So the time sample and topography similarity over samples are very important for me. About what you mentioned " representational similarity analysis" or " threshold free cluster analysis ", Could you please introduce reference study for ERPs, EEG or similar signals?
Sincerely,
Reza
Dear Reza
We often calculate the coherence in their phase and wavelet-based methods to study dynamical interrelations between brains signals on rest-state EEG data. Bicoherence is a method to measure the related between two frequency on single channels or different channels. I am not sure whether bicoherence can be used to calculated ERP data. you can search some article to find that. By the way , we often use wavelet bicoherence to get a similarity matrix. the picture is a example of bicoherence matrix.
Hope a little help for you.
best
Zhou Tianyi
ERPs can be subjected to overlapping-window frequency analysis to generate ERD/ERS, event related de/synchronization -- and bicoherence can be performed on ERD/ERSs. As can bimodulation, automodulation, and numerous others
Dear Tianyi,
Thank you so much for considering!
That's very good point to use biocoherency for EEG data to find bicoherence similarity matrix based on frequency.
Now for ERP data, the pattern based clustering is more reasonable for example I used BiClustering and pClustering for that purpose. Here is a good reference for my motivation,
https://dl.acm.org/citation.cfm?doid=564691.564737
Sincerely,
Reza
EEG is numbers by the pound so endless elements to cluster in time and frequency. Clustering sounds like information compression techniques where the findings of each cluster-compression tells something useful in the market or whatever data set was used.
Thanks everyone!
As a short conclusion, for ERP data we could consider a topography as a pattern, so we may use pattern based clustering such as Biclustering, pClustering(I added a good reference in previous answer) or correlation based clustering algorithms such as; K-means and Hierarchical Clustering.
For EEG we may use Bicoherency in terms of finding similarity between frequency bins.
Sincerely,
Reza
One has to be careful / cautious about using k-means and correlation; k-means is meant to work the Euclidean distance (the update rule of the centroids).
The other alternative, is the similar, K-Medoids.
- Badges
- Science topic