Why did we use the threshold of 2.9
gowthambalachandhiran opened this issue · comments
I think, this reconstruction distance based method for fault conclusion valid. should take
the count based on distance linspace counts change.
When the value mass blow change into small, this is the character for fault.
Not simple setting to one num, but can change with precision recall you need.
one facet can defined by precision and recall in valid set, the other facet may be
defined by diff between linspace range num count, use the neighborhood of monotone change point
choose as the threshold.
And the Second-order derivative of linspace range num count may be a prefered a criterion for early-stop of autoencoder, the train loss for generalize the model but the Second-order derivative of linspace range num may measure the discrimination between fraud and normal points.
This is my suppose, Dose there some criterions for convergence of unsupervised learning for fraud detection ?
If you know, Please tell me.