3-Dimensional Matrix for Machine Learning/Logistic Regression

Question

I am wondering if anyone has tried using a 3-dimensional matrix for logistic regression? I am working on a project that involves audio where for a matrix X, the n dimension is the features for each audio sample and the m dimension is the number of audio files I am comparing to each other. The problem I'm facing is that for each audio file, there are numerous audio samples (every 44100 samples, I will take as "1 sample"), each of which includes each different feature. Instead of just taking the average of each feature for every audio sample, I was considering adding a third dimension to the matrix as described above that is the audio samples. The thing is, I don't know how this will hold up in logistic regression/what I can expect my output variables y to be. Does anyone have any experience with this?

The result I am looking for is a way to take all of the features from each audio sample and compare several audio files to come up with an algorithm that can determine a "good" audio mix from a "bad" audio mix, so I believe (though I am not certain) that I am only dealing with two output classes here (good vs. bad).

Thanks for the advice

Answer 1

Since you're dealing with audio input, you should probably use a model that was designed to handle sequences (audio=sequence of frames).

I would expect that (say) training an LSTM on audio signal would perform at least as well as logistic regression trained on the mean frame. However, it would likely take much longer to train.

You might want to start with one of the github projects listed under: https://github.com/topics/audio-classification (I've never used any of these).