[x] Grid search for initialization hyperparameters: including number of folds
[ ] check variance in the results for different initializations (different initializations * different folds)
[ ] Compute model performance (accuracy metric)
[ ] find smallest K that is not significantly different from K with max LL
[ ] Gap G statistic
[ ] look at how states change for K and K+1 (is it one state which splits or does everything change?)
[ ] Look at the trial epoch state statistics across folds - are results very different across folds?
[ ] Check values on test set to understand why model overfits so fast
[x] Should select stickiness parameter based on cross-validation
[ ] recovery procedure: can define states (labeled) with respect to the trial epochs; can then get emission matrices from there; do we recover the different time intervals?
[ ] look at LL on training set
[ ] find a way to deal with best params on the edge
[x] Implement Bayes factor
[x] Grid search for initialization hyperparameters: including number of folds
[ ] check variance in the results for different initializations (different initializations * different folds)
[ ] Compute model performance (accuracy metric)
[ ] find smallest K that is not significantly different from K with max LL
[ ] Gap G statistic
[ ] look at how states change for K and K+1 (is it one state which splits or does everything change?)
[ ] Look at the trial epoch state statistics across folds - are results very different across folds?
[ ] Check values on test set to understand why model overfits so fast
[x] Should select stickiness parameter based on cross-validation
[ ] recovery procedure: can define states (labeled) with respect to the trial epochs; can then get emission matrices from there; do we recover the different time intervals?
[ ] look at LL on training set
[ ] find a way to deal with best params on the edge