Choosing Histogram Bins for Time-Indexed Events
A histogram of timestamps estimates an event rate, and its error has a minimum in the bin width. See how to estimate that minimum from the counts, and when the method fails.
-
-
Survival Analysis
Some rows have not happened yet, and both obvious fixes are wrong in the same direction.
-
Recommender Systems
The evaluation protocol moves the number more than the model does.
-
Time-Series Forecasting
The rows are not exchangeable and the baseline is strong.
-
Semi-Supervised and Active Learning
Unlabeled data is free; the assumption it comes with is not.