I have about 50k devices, which either fail (0) or connect (1). I have data for the past 300 days for all of them. I'm trying to predict when a device will next fail. The devices are all more or less the same type.
So, I have a matrix like Device 1: [0, 1, 1, 0, 0, ...] Device 2: [1, 0, 0, 0, 0, ....] Device 3: [1, 0, 0, 0, 1, ....] ...
And trying to predict the next column.
So far, I've tried XGBoost / deep gated recurrent nets. I'm wondering if there's a better approach.
I think the data from all 50k devices should be pooled in some way, then the final prediction should be tailored to the characteristics of the individual device. Not sure how to do this statistically.