Hey everyone,
currently, I am working with Gaussian process latent variable based models. In literature, the model likelihood is discussed for model selection.
Unfortunately, this does not work for my application. Currently I am using the log likelihood and the reconstruction error. The model likelihood increases and the reconstruction error decreases with increasing dimension/inducing points. BIC doesn't make sense in this context (and behave similar...).
Are there better parameters for model selection?
All the best,
Will
Refer this link
http://proceedings.mlr.press/v9/titsias10a/titsias10a.pdf
https://www.ias.informatik.tu-darmstadt.de/uploads/Teaching/RobotLearningSeminar/Goppalan_RLS_2012.pdf
Thank you Karthika .P ,
I know these publications - I already implemented an ARD based model selection approach. But I have following problems:
- I can choose the dimension of a given model using the ARD values
- The number of inducing points is still unsolved
- An analysis of reconstruction error and model likelihood shows, that more dimensions and inducing points will leads to better results (this is a good thing)
- For my application, I want to find the information content in the latent space. I could do this using supervised learning afterwards (e.g.: Boruta [1]), but I do not want to do that because I am interested in the information content of the latent space itself without prior knowledge (such as labels).
- I do not want to use approaches based on the fitted GPLVM model.
So ARD, log. likelihood and reconstruction error are not appropriate in my context.
[1] https://www.kaggle.com/cdupin/using-boruta-feature-selection-wtih-random-forest