Is there a way to force certain matches in pot?
Hello again, Thanks for addressing my questions about using pot for large matrices, it works beautifully! Now, I have a situation in which I want to force certain matches to happen (namely, I want to ensure that certain positions in the output matrix have a specific, non-zero value). Do you think there's a way to use pot for this case? Thanks, -Vanessa
Hello Vanessa, It's a nice problem you have. We have a simple way to forbid mass on some links explained in the semi-supervised section 4.2 of http://remi.flamary.com/biblio/courty2016optimal.pdf Basically we set the loss to a very large value instead of the distance which will forbid some mass transfer. If you want to force a given source to go to a given target you can do something similar using very large values on the line/column except when they cross where you use 0. Another approach would be to have a G=Gprior+Gopt where Gprior is fixed a priori and you want to find the remaining optimal Gopt. I would personally remove the marginals of Gprior to a and b then normalize again to 1 and solve the problem to find the Gopt and inverse the normalization to compute the final G with G=Gprior+Gopt. The proposal above do not force the mass value but force a lower bound on the mass transfer in a given link which might be enough for you. Rémi Le 29/08/2017 à 22:45, Vanessa Ortiz a écrit :
-- Rémi Flamary Web: http://remi.flamary.com Tel: +33 (0)4 92 07 63 80 Laboratoire Lagrange, UMR CNRS 7293 Observatoire de la Côte d'Azur Université Côte d'Azur
Hello Vanessa, It's a nice problem you have. We have a simple way to forbid mass on some links explained in the semi-supervised section 4.2 of http://remi.flamary.com/biblio/courty2016optimal.pdf Basically we set the loss to a very large value instead of the distance which will forbid some mass transfer. If you want to force a given source to go to a given target you can do something similar using very large values on the line/column except when they cross where you use 0. Another approach would be to have a G=Gprior+Gopt where Gprior is fixed a priori and you want to find the remaining optimal Gopt. I would personally remove the marginals of Gprior to a and b then normalize again to 1 and solve the problem to find the Gopt and inverse the normalization to compute the final G with G=Gprior+Gopt. The proposal above do not force the mass value but force a lower bound on the mass transfer in a given link which might be enough for you. Rémi Le 29/08/2017 à 22:45, Vanessa Ortiz a écrit :
-- Rémi Flamary Web: http://remi.flamary.com Tel: +33 (0)4 92 07 63 80 Laboratoire Lagrange, UMR CNRS 7293 Observatoire de la Côte d'Azur Université Côte d'Azur
participants (2)
-
Rémi Flamary -
Vanessa Ortiz