How to use a pre-trained model to predict on new data

#1
I am trying to look for examples where a pre trained model on one kind of dataset can be used to predict a new kind dataset and I would like to explored methods that do not use deep learning. I understand that it is possible to use a pre-trained model as an input for a new model but I am not sure how exactly this can be done from a statistical point of view.
 

hlsmith

Not a robit
#2
This depends on your purpose. I can train a model using a dataset and cross-validation. Easy enough, then apply that model to a new dataset. So just use those selected features (variables). Another option is to do this but don't use cross-validation but use a random holdout set to select the model. Or are you writing about something different? What is the purpose of you wanting to do this and is the new dataset a random holdout or say the next year's data, etc.?
 
#3
my purpose is about using a pre trained model for a dataset of a different domain. it is usually called transfer learning but using the term transfer learning is usually paired with deep learning. my intention is not to use a black box model but to find a more explainable method for this.
 

hlsmith

Not a robit
#4
If it is a different domain, meaning the model was built on a sample from a population and you want to use that model on a sample from different population, there is a term called transportability, which functions to transport results to a different 'domain'. I haven't done this before, but there have been many paper on this over the past year. There is a person at Purdue University who has been working in this area:

https://causalai.net/
 
#5
thank you very much. I will try using this keyword tranportability in my search as well. because at this point using transfer learning as a keyword often points to examples using deep learning which I would like to avoid for my case.
 

hlsmith

Not a robit
#6
Good. The example I always hear in regards transportability is say you conduct a study or institute a policy in Los Angeles California. Now you wonder what the impact of it would be in New York, New York. Well the two samples have different characteristics and NYC my have a secondary variable that is an effect modifier, How do you reweight results to "transport" them to a sample from a different population, -transportability. This process also uses directed acyclic graphs in order to confirm transportability and what may need to be addressed.
 
#8
Yes.. i would think that my use case falls in the same category as well but I am just amazed at how many terms are being used to almost depict the same idea.. within this scope itself i could use as many as 10 different search terms that points towards the same type of task. anyways i still need to also ensure that the results can indeed be transportable because if the two domains do not have any relations whatsoever then it will also result in something called negative transfer learning therefore a good understanding about the source domain and the target domain needs to be first addressed before any kind of transfer learning is executed.