Tuning Binary PCA using Generalized Low Rank Model

Question

I am using h2o Generalized Low Rank Models for tuning a binary PCA (Logistic loss + Quadratic regularization) to perform a recommender system task. I am using the H2O infrastructure using the GLRM approach. Below the code where a model is run.

bestGlrm<-h2o.glrm(training_frame = train.h2o.wmiss, cols = glrm_cols, k = best_k, validation_frame = train.h2o.nomiss, seed = 1,
     loss_by_col=rep("Logistic",ncols), regularization_x = "Quadratic", regularization_y = "Quadratic", gamma_x = best_gamma_x, 
     gamma_y = best_gamma_y, transform = "NONE", impute_original = FALSE,model_id = "bestGlrm")

Key tuning parameters are: gamma_y, gamma_x (regularizers) and k. Regarding the gamma x and y, the parameter space lies from 0 to $\inf$. I would like to know whether there is any guidance to select the gamma range on data.

Answer 1

I would suggest that you use the gridSearch on GLRM and perform gridsearch over gamma_y and gamma_x. The runit_glrm_grid_iris.R in h2o-3/h2o-r/tests/testdir_algos/glrm show you how to do gridsearch.

My proposal will be to first search gamma_x and gamma_y over a wide range with wide intervals. As you narrow down the range where the metric looks good, you can then tighten your intervals. For example, you can start with 0.0001 to 1000 with step of 50. From your gridsearch result, it shows that the best result is between 50 and 200, then, set your next gridsearch from 50 to 200 but with a step size of 5.