I have a following classification problem with somewhere around 2000 examples.

For each example, there is a feature vector $\bf{x}$ of size $N$ and a label vector $\bf{y}$ of size $L$. Each label entry can take 3 values from a label-set $S\in\{-1,0,1\}$. The label vector has some structure to it such as there could be up to 3 entries with -1 and maximum one entry with 1 etc etc.

so for example y could be [-1, -1, 0, 1, 0, 0] for L=6

How to approach this kind of problem? Any hints/suggestions?

  • $\begingroup$ You can use the simple binary relevance method that is also used in normal multi-label problems. $\endgroup$
    – PhilippPro
    Dec 10 '17 at 8:36

I don't know if this is a good idea, but if you regard each entry in your vector as a variable -- which I assume it actually is -- say A, B, C, ... F in your example, you can encode the values as M1, 0 and P1 (for -1, 0, and +1, respectively) and combine the two to create actual labels. Your example's labels would thus be: [AM1, BM1, C0, DP1, E0, F1].

I believe this would work with standard multi-label algorithms. The problem would be that prediction would not constrain the labels to follow your other rules (3 minus, 1 plus, etc).

But I don't think that thinking of your S as labels is going to work in the multi-label literature/paradigm since they are not really labels.


Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.