search

robobenklein / lspz

0 stars 0 forks source link

Relative comparisons to absolute value approximation #2

Open robobenklein opened 1 year ago

robobenklein commented 1 year ago

Inputs: a series of keys A vs B, ranging from -1 to 1, specifying A or B to be strongly/weakly greater/lesser than the other.

Outputs: the most optimal assignment of values to keys where the comparisons are most strongly represented in the sorted ordering

Given the size of my own library, there are about 35k possible 30-second samplings, and almost 5k tracks, so the performance should be good in order to be useful.

robobenklein commented 1 year ago

converting individual comparisons based on A>B likelyhood to a ranking: https://stackoverflow.com/a/17701105/2375851

robobenklein commented 1 year ago

ELO estimation algorithms could be used to fit individual items likelyhood to be greater to a range of scores: https://chess.stackexchange.com/questions/37352/what-sets-the-absolute-value-of-players-elo-rating

robobenklein commented 1 year ago

https://core.ac.uk/download/pdf/226942134.pdf

From pairwise comparisons and rating to a unified quality scale

robobenklein commented 1 year ago

Using Pairwise Comparisons to Determine Consumer Preferences in Hotel Selection https://www.mdpi.com/2227-7390/10/5/730

super math-heavy

robobenklein commented 1 year ago

Inverse: https://en.wikipedia.org/wiki/Bradley%E2%80%93Terry_model

robobenklein commented 1 year ago

Gradient descent: while initially sounds like a valid option, the size of the set of possible comparisons (N samples compared to N samples) makes the computation expensive and the iterations very slow.

parallelization of similar algorithms in deep learning could probably be useful to speed this up and perhaps narrow the number of samples needed in each comparison

https://shashank-ojha.github.io/ParallelGradientDescent/ (conclusion: GPU-sized parallelism didn't help, 8-16 cpu cores was their limit for speedup)