Please, how to calculate the number of model parameters of the random forest model to determine the Akaike Information Criterion (AIC) using this following expression:
AIC = n * log(MSE) + 2 * p
with n is the number of training samples and p is the number of model parameters.
For understanding Machine Learning or Data Mining techniques, I suggest you to follow the attached book and http://myweb.sabanciuniv.edu/rdehkharghani/files/2016/02/The-Morgan-Kaufmann-Series-in-Data-Management-Systems-Jiawei-Han-Micheline-Kamber-Jian-Pei-Data-Mining.-Concepts-and-Techniques-3rd-Edition-Morgan-Kaufmann-2011.pdf
Dear Hamza,
random forest is totally different algorithm and does not work with likelihoods like regression and n is also not that clear. So, AIC will not work here. For what exactly do you need this? Maybe there is another solution for your requirement
Information criteria are used for model selection: in the case of machine learning this is done using resampling methods.
The numbers to be plugging into the formula make little sense for a random forest.
If you really need to select specific values for the random forest, use cross validation
This a comment of a reviewer who asks me to use the AIC to compare machine learning algorithms such as ANN, SVR, GB and RF. He recommends me an article in which they calculate the AIC of the ANN using the number of parameters which is the number of weights and biases. However, I don't know how to calculate the AIC for other algorithms.
In fact, I divided data into training and testing sets using cross-validation technique. I used the testing set to compare machine learning algorithms using the RMSE and MAE metrics. However, the reviewer wanted to perform the comparison in terms of accuracy and simplicity using AIC indicator.
Hi Hamza Ali-Ou-Salah , you should have put that information in the answer in the first place, in order to help the readers help you in return.
Sometimes reviewers have absurd requests
You can tell the reviewers that Information Criteria for machine learning are not applicable, because the number of parameters is not at all clear: Information Criteria are meant to penalize the likelihood for model complexity; but
what is the number of parameters in a RF? The number of splits, the number of trees, the depth of the tree? That is not clear
Machine Learning models have to be compared out of sample;
- Badges
- Science topic