investigate weighting samples within the NaiveBayes approximation based on time-since-observation
We can implement schemes like an
- FIR filter
- remember time stamps of each sample and only keep the recent ones
- may be expensive/require re-computation of the entire histogram during each call to evaluate
- IIR filter
- update the entire histogram in some manner.
There should naturally be a parameter that controls how long the algorithm remembers features. For example, if the IIR approach defines hist_new = 0.9*hist_old + data_new
in some sense, then it will gradually forget about old data as the histogram is updated repeatedly.