Take into account uncertainty in ground-truth in a classification I need to evaluate the accuracy of a prediction model (binary classification), given that I have some kind of uncertainty on the ground truth measurement.
The model predicts the occurrence of an event at time T, T+5mn, T+10mn, etc.
Unfortunately, the measurements are less frequent, and not aligned. For instance, it can say that an event occurred between T+2 and T+17 (15mn period, not aligned).
If both prediction and measures are negative on a period, then I have true negatives.
But some configurations have uncertainties: if the prediction is negative at T+5,T+10,T+15, but the measurement in the range [T+2,T+17] is positive, I can deduce that at least one of the 3 predictions is a false negative, but maybe not all.
Is there a way to model this uncertainty in the evaluation of the prediction ?
A basic way would be to count the best-case and worst-case false negatives and get a range in the end (and the same for positives). Any better way ?
 A: It sounds like the uncertainty in your ground truth stems primarily from uncertainty about the true duration and intensity of the phenomena's presence: when the sensors show positive, it's not clear whether that means the phenomena has been present at lower intensities for a longer spans or at higher intensities for shorter spans.
Given that uncertainty, one option would be to systematically evaluate the model's accuracy over a broad range of scenarios, including some where the ground truth reflects short bursts of high intensity and others where the ground truth reflects longer periods of lower intensity.  These alternatives could be captured by smoothing the observed binary sensor values over shorter or longer windows of time.  Rolling averages provide one simple way to accomplish such smoothing.  If the temporal window over which one averages remains small (zero smoothing being the smallest), then one effectively assumes high intensity inputs over the entire [T+2, T+17] interval.  If, on the other hand, the window over which one averages becomes as long or longer than the interval itself, then one is assuming that the phenomena operated at lower intensity but over a longer period of time.
I don't know how much information you have about the "external factors" influencing the intensity of the phenomena and its likely durations, so I can't say what might constitute a reasonable range of window sizes to try.  But examining the minimum and maximum durations during which the sensors showed positives might help.
In any event, I think you might be able to account for your uncertainty about the true inputs by evaluating your model's accuracy under a variety of plausible duration/intensity scenarios, which could, in turn, provide you with a reasonable range of possible accuracies.
