Imbalanced classification

[logictensornetworks]

Imbalanced classification

[sbadredd]

Hi @mpourvali, thank you for the feedback!

define variables for which in each batch they might be empty?

This is the main challenge, right?

I should add support by ltn for cases where a variable is empty. For example, Forall(x_empty,P(x_empty)) would be accepted and would return 1. Exists(x_empty,P(x_empty)) would return 0.

Then you could write the same axioms at every batch, even if some are not used. Would that be a good solution?

[mpourvali]

Hi @sbadredd , thank you for reply!

This is the main challenge, right?

yes, its's the main challenge,

Would that be a good solution?

it sounds a good solution for satisfiability, and then can I know what P(x_empty) would return?
can we initialize variables?

[sbadredd]

I expect that P(x_empty) would return an empty tensor. Every formula connected to it, e.g. And(P(x_empty), P(x_nonempty)) would also return an empty tensor. In that scenario the system would not throw exceptions if you initialize a variable with an empty domain

One potential issue is that empty domains can lead to semantical problems (e.g., some axioms for the t-norms and quantifiers not holding anymore). I will discuss this with other collaborators before implementing it

In the meantime, can you consider these temporary solutions?

axioms = []
if tf.not_equal(tf.size(x.tensor), 0):               # Only if the tensor of variable x is not empty...
    axioms.append(...)                                      # ... add the axiom about x