machine learning - Why does RSquared increase with # of folds in k-fold cross validation? -

i'm tuning model using k-fold cross-validation , noticed rsquared accuracy appears improve number of folds -- e.g. higher rsquared value when using 30 folds compared using 10 folds.

two questions hoping insight on:

1. why occur?
2. is there reason believe rsquared k=10 better estimate of model accuracy using k=30? or both unrelated future error rate can expect on unseen test set?

here's simple example of effect i'm referring to:

############### k = 10 ##################### > data(iris) > train_control <- traincontrol(method="repeatedcv", number=10, repeats=3) > train(sepal.length~.,data=iris,trcontrol=train_control,method="rf",metric="rsquared")  random forest   150 samples 4 predictor  no pre-processing resampling: cross-validated (10 fold, repeated 3 times)  summary of sample sizes: 137, 135, 134, 134, 135, 136, ...  resampling results across tuning parameters:  mtry  rmse       rsquared   rmse sd     rsquared sd 2     0.3381065  0.8404534  0.07692415  0.07583768  3     0.3247406  0.8502577  0.07311807  0.07326181  5     0.3228651  0.8517740  0.07213958  0.07315720   ############### k = 30 ##################### > data(iris) > train_control <- traincontrol(method="repeatedcv", number=30, repeats=3) > train(sepal.length~.,data=iris,trcontrol=train_control,method="rf",metric="rsquared")  random forest   150 samples 4 predictor  no pre-processing resampling: cross-validated (30 fold, repeated 3 times)  summary of sample sizes: 143, 145, 146, 144, 145, 144, ...  resampling results across tuning parameters:  mtry  rmse       rsquared   rmse sd     rsquared sd 2     0.3238545  0.8580474  0.10327919  0.1352787   3     0.3119541  0.8679321  0.09734168  0.1236307   5     0.3109572  0.8717550  0.09727307  0.1123173   

bigger number of folds - bigger training set , smaller testing one. observe best result fo loo (n-fold n training samples) , worst k=2. there no 1 answer generic question how many folds use, solely depends on dataset. furthermore if there underlying relation between datapoints (for example come time series) important how set divided.