hoxo-m / densratio Goto Github PK

The uLSIF algorithm find optimal lambda and sigma parameters using grid search.
Instead of grid search, how about use Bayesian optimization?

First off, thanks for this very useful package! The interface is smooth and the implementation is quite fast, even for large problems.

I've noticed that some papers make use of regularization penalties when fitting weights for KLIEP (see, e.g., §3.2 of this review). I wonder if it would be possible to incorporate L1, L2, and/or elastic net regularization into the KLIEP optimization? Something similar appears to be implemented for uLSIF.

http://www.kurims.kyoto-u.ac.jp/~kyodo/kokyuroku/contents/pdf/1703-02.pdf
によれば最小二乗の場合はAICもモデル選択に有効のようですが、AICは実装されていないのでしょうか or AICをdensratioの結果に対して計算するためには何を尤度と考えるのが適切でしょうか？

This line probably computes the Gaussian kernel.
If so, I think the Euclidean distance should be squared.

There are some bugs in the script KLIEP.R. One is that in this sentence if(verbose) message("Searching optimal sigma and lambda...") we only optimize sigma without lambda. A serious one is in the function compute_density_ratio, this function use the train set to predict the density ratio of a new sample, which results in the mismarch of the dimension of the predicted value and the new sample.