Let's say you have multiple sources of observation weights for a dataset. For example, you have a $[0,1]$ weight coming from the label's certainty ($w_c$) and another one coming from its recency ($w_t$). What would be the best way to combine those into a single number to be used when training a classifier?

Some options I have considered:

  • Product: $w_cw_t$
  • Arithmetic mean: $\frac{w_c+w_t}{2}$
  • Geometric mean: $\sqrt{w_cw_t}$
  • Harmonic mean: $\frac{2w_cw_t}{w_c+w_t}$

All of the above make (intuitive) sense on some level. Any ideas as to what is best practice or at least some pointers to literature on the topic?


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.