I am trying to predict failures at different nodes on a line. Each node has different weather features and hardware/configuration features. For a little under half of the historical failures I have, I know the precise location/exact node where the failure happened, however for the other half I only know a range of nodes the failure could have happened on. My current approach is to "smear" a 100% probability across N nodes that the failure could have happened on. e.g. I know a failure happened between nodes A and D therefore assign A, B, C, and D the value 0.25.

I am currently treating my model as a regression on the probability at each node in keras with a single node output layer but the performance is really poor. For other problems like this, historically I've had plenty of data and more exact locations so I could treat it as a strict binary classification however I've never had to deal with classifications that are "partial" so-to-speak. How would you handle this problem? Is a regression of the probability the best approach? I've had the thought of assigning all possible nodes full 1's instead of fractions and treating it as a standard binary classification but I'm worried that will greatly over represent the unknown location failures.


Your Answer

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

Browse other questions tagged or ask your own question.