Classification Learner and Regression Learner: Please help!

Hello Everyone,

I am trying to use the classification learner app. I have 2 sets of data, one considered "good" and one "bad". I am trying to create a model (either an SVM or regression). But I do not know how and I keep getting poor accuracy. The data is signals. Each set of signal data is around 500 signals. Each one of these signals is represented by a row of 2000 numbers. How can I train a model to tell the difference between good and bad signals and use it to filter through other signal data?

[Walter Roberson]

Okay, 500 signals, 2000 samples per signal. Are the 500 signals to be considered independent, like 500 examples of signals, or are they 500 correlated channels (for example 500 microphones at different positions being used to figure out distances to signal sources by time delays) ?

Are the individual signals time courses in reaction to an event, or continuous repetitive signals, or non-repeative audio (e.g., speech or music) ?

[lus31]

Thank you!

500 independent signals. (They are obviously similar, but correspond to different points in time/different locations)

The signals are ultrasound. So at time t/location x, a signal is sent through the medium and then the response is recorded. The signal sent through the medium each time is the same, but the response is not neccessarily, hench why I want to separate good from bad.

[Walter Roberson]

You should probably do some feature extraction and train based on the extracted features... but it might not be easy to figure out what to extract.

SVM and regression are not suited for the general case of just throwing in some data and hope that something interesting shows up: you would want Deep Learning techniques for that. Or at the very least a time-series network with more standard NN techniques.

[lus31]

So is there a way to do feature extraction with Matlab?

I do not really want to just choose random signals to use as a baseline. The good signals are very similar to each other, but none are exactly the same.

Ideally, I just wanted to use those 500 good signals to train a model of what a good signal would look like.

Thank you!

[Walter Roberson]

What if you took the mean of the good signals and calculated the correlation to the mean? The bad signals might have a quite different correlation.

[lus31]

Thats not a bad idea, but wouln't that mean I am only basing my model on one feature?

[Walter Roberson]

Yes, but it might be enough.

(Remember, we haven't seen any samples of good and poor signals to be able to guess what other features might work.)

[lus31]

Okay thank you for your help, but one last question..

So I create this mean signal, should I just load that into the classifcation learner and then output the model?

[Walter Roberson]

Calculate the correlation between the mean of the good signals and each individual signal, and use that correlation as the input.