search

weecology / LDATS

Latent Dirichlet Allocation coupled with Bayesian Time Series analyses
https://weecology.github.io/LDATS
Other
25 stars 5 forks source link

Flexibility in number of parameters in the changepoint model #128

Open ha0ye opened 5 years ago

ha0ye commented 5 years ago

My understanding of the issues around selection of the # of changepoints (weecology/MATSS-LDATS#35) is that the number of parameters grows ~ (# of covariates) x (# of community groups) x (# of changepoints).

As @juniperlsimonis mentioned, this is partly because the parameters are independent, as opposed to being drawn from a distribution with hyperparameters, and that supporting the latter may help with the issues with explosion in number of parameters.

Another possibility might be to not need independent values for all the regression coefficients across every changepoint for every community group. If we can flexibly have intercepts and slopes as sometimes unchanging across changepoints, that could reduce # of parameters (especially if changepoints are only affecting some of the community groups, but not others).

(filed under "thoughts on adding flexibility to the model" for future versions of LDATS)

juniperlsimonis commented 5 years ago

totally good idea on that approach. it'll definitely require more computation, tho (like most things with this method), so is another benefit of having more efficient coding for the computation under the hood.