Weighted harmonic mean of coherence and exclusivity
Source:R/model_selection.R
coherence_exclusivity_score.RdA single score to rank models on the coherence-exclusivity trade-off: $$H = \left(\frac{w_c}{c} + \frac{w_e}{e}\right)^{-1},$$ with \(c\) and \(e\) the (rescaled) coherence and exclusivity and \(w_c + w_e = 1\). A harmonic mean penalises models that are poor on one criterion.
Usage
coherence_exclusivity_score(
coherence,
exclusivity,
weights = c(0.4, 0.6),
rescale = TRUE,
floor = 0.01,
...
)Arguments
- coherence, exclusivity
Numeric vectors.
- weights
Numeric vector of length 2, weights of coherence and exclusivity (normalised to sum to 1).
- rescale
Logical: rescale each criterion to [0, 1] over the models given (min-max) before averaging. Needed for UMass coherence, which is negative (a harmonic mean of a negative and a positive number has no meaning). The score then ranks the models given and is not comparable across grids.
- floor
Lower bound of rescaled values, so that the worst model on one criterion does not get a score of 0 whatever its other value.
- ...
Not used: an argument renamed in the development version of mintR gives an error naming its replacement.