My first data profile has 3 dimensions; atom pattern (I think it is similar to frequency), time, and signal intensity. I can use supervised NMF for source separation straightforwardly. Because the atom pattern can be overlap, the first result is just acceptable.

I can retrieve the second data profile as a Boolean vector. At specific time point, some atom pattern has another underlying pattern which can be describe with a Boolean vector. Are there some good ways to incorporate another dimension of information as a boolean vector into NMF model? I try to find a similar data in audio processing, but not yet found. Any reference or idea are greatly appreciate.

enter image description here

  • $\begingroup$ I don't quite understand what your ultimate goal is. Could you explain: 1. what are your sources? 2. why is the first result just acceptable and how is this related to the atom patterns overlapping? 3. how do you retrieve the second data profile? 4. what is the second data profile? 5. how are the atom patterns related to another underlying pattern that can be described with a Boolean vector? $\endgroup$
    – applesoup
    Jan 2, 2019 at 15:15
  • $\begingroup$ In addition to my previous comment: Can you formulate a cost function that you want to minimize? If so, it is probably much easier for people to help as the complete optimization problem is described by the cost function. $\endgroup$
    – applesoup
    Jan 2, 2019 at 15:16
  • 1
    $\begingroup$ @applesoup I fixed as you suggested and posted here dsp.stackexchange.com/questions/54578/… . I would be very appreciate if you may take a look. Thank you very much. $\endgroup$
    – Jan
    Jan 3, 2019 at 12:22


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.