The AUC in Task 1

[adriaciurana]

Hello, we have read task 1 and we have to calculate the AUC. But the problem is that the curve between precision and recall needs a threshold to be created and in this task 1 does not specify what's this threshold (Alpha, rho, etcetera). Thanks!

[Guim3]

Also, can we calculate the F1-Score instead of the AUC? The first one is easier to compute and is also a good measure to evaluate a model.

[jrhupc]

Hello,

The idea in task 1 is to use the best configuration (the one with the best AUC) that you obtained in week 2 and improve that with a morphology post-filtering.

It is true that you could use F1-Score but in order to be able to compare results between groups please keep using AUC for evaluation.

Hope things are clearer now. Best regards,

[priba]

One last thing. We understand that we should take the best model obtained in the previous block, in our case, one gaussian recursive with a fixed alpha and rho for each sequence. Hence, we have two parameters alpha and rho. Thus, we have to compute all the curve in 3D with both parameters and compute the area under the curve or we should fix one of these parameters?

From the last class, we thought that we have to apply the morphology directly to these model with the fixed parameters. Finally, for every connectivity in the task 1 and every number of pixels in task 2 we have to move the threshold for the parameters and compute the AUC?

Thank you.

[jrhupc]

Hello,

I would recommend to fix rho and use alpha to compute AUC in Task 1. For task 2, fix connectivity and rho and compute AUC for each value P and plot a graph to estimate the optimal value of P.

In any case, it would also be good to include the F1-Scores along the tasks to see if results are similar to that with AUC.

Regards,