Great work and interesting project to get into R. I get this error but can't see why I get this error "Final_Model.RDS', probable reason 'No such file or directory'"
Hey when I run the Modeling.Rmd file, at line 646 I get an error for this keras_tune % mutate(weights = map(iteration,get_dirichlet) ,res = map(weights, tune_keras_weights)) with the error being "matrix type cannot be converted to python (only integer, numeric, complex, logical, and character matrices can be converted" Any advice on how to resolve this?
I think since it mentions python it could be a reticulate error. So you might need to specify your python environment using use_python() or use_virtualenv(). It could also be an error with the loss weights needing to be in a different format.
Interesting video, too bad the rds files were corrupted. With feature engineering, I am never sure whether the features I am generating would automatically be constructed by the model, if they were indeed important. Especially, relatively simple combinations of input features (+, -, *,/). I've noticed in the past that I generate a new input feature for the model that I think is really clever and it shows up near the top of the vip chart but the overall model performance is not significantly better than without the new feature. Do you know of any "rule of thumb" guidance for which combinations are typically not useful? I suppose it could be different depending upon the model type.
I'm not sure if there is any one rule of thumb for determining which features are useful. The fix for the scenario you mentioned is something that depends on the data structure (sample size, dimensionality, model type, etc.). I think for relatively simple combinations, it is probably better to remove them but I wouldn't use it as a definitive rule especially if the EDA shows a specific relationship. This is not shown in the video, but I usually try a decent amount of methods such as tree-based models, lasso models, recursive feature elimination, VIP, and PCA to regularize my models. For me, Lasso models and VIP are probably the most used in my own work but all of the methods could be used. I would like to do more regularization on the individual component models but to be honest I ran out of time. Maybe I'll make a video exploring different ways to select features for this project.