tune_BHPMF: Tune latent factor size in BHPMF.

In BHPMF: Uncertainty Quantified Matrix Completion using Bayesian Hierarchical Matrix Factorization

Description

Tune the latent factor size (k) in BHPMF using cross validation.

Usage

TuneBhpmf(X, hierarchy.info, prediction.level, used.num.hierarchy.levels,
		  	 	num.folds=10, num.samples=1000, burn=100, gaps=2, tmp.dir,
		  	 	verbose=FALSE)

Arguments

X	The data matrix (including missing values). Rows are the observations, columns are the features. Missing values must be filled by NA.
hierarchy.info	A matrix containing the hierarchical information in the same order as matrix X. Data should be ordered in descending order (the lowest level being in the last column). The first column is the observation index.
prediction.level	The level at which gaps are filled. For example if set to 4, gaps are filled at the 3rd highest hierarchy level. The default value is at the observation level (filling gaps/missing values at the observation level).
used.num.hierarchy.levels	Number of hierarchy levels used for gap filling. For example, if set to 2, only information from the first and second levels of the hierarchy are used. The default value equals the total number of hierarchy levels.
num.folds	Number of cross validation folds for tuning. The default value is 10 if tuning set to TRUE.
num.samples	This is aparameter for Gibbs sampling: total number of generated samples at each fold using gibbs sampling. Note: this is not the effective number of samples. The default value is 1000.
burn	This is aparameter for Gibbs sampling: burn in period. Number of initially sampled parameters discarded. The default value is 200. Should be smaller than num.samples.
gaps	This is aparameter for Gibbs sampling: gap between sampled parameters kept. The default value is 2.
tmp.dir	A temporary directory used to save preprocessing files. If not provided, a tmp directory in /R/tmp will be created. If provided, each time a function from this package is called, the preprocessing files saved in this directory will be used. Helps to avoid having to re-run preprocessing functions for the same dataset. WARNING: This directory should be empty for the first time call on a new dataset!
verbose	If TRUE, progress of the sampler details are printed to the screen. Otherwise, nothing is printed to the screen. Default is FALSE.

Value

A list with the following tags:

best.number.latent.features	The best value for latent vectors length
min.rmse	The minimum RMSE for the latent vectors of size best.number.latent.features

Author(s)

Farideh Fazayeli [e-mail]

References

F. Fazayeli, A. Banerjee, J. Kattge, F. Schrodt, P.B. Reich (2014) Uncertainty quantified matrix completion using Bayesian Hierarchical Matrix factorization. International Conference on Machine Learning and Applications (ICMLA)

F. Schrodt, J. Kattge, H. Shan, F. Fazayeli, A. Karpatne, A. Banerjee, M. Reichstein, M. Boenisch, S. Diaz, J. Dickie, A. Gillison, V. Kumar, S. Lavorel, P.W. Leadley, C. Wirth, I. Wright, S.J. Wright, P.B. Reich (2015) HPMF - a hierarchical Bayesian approach to gap-filling and trait prediction for macroecology and functional biogeography. Global Ecology and Biogeography 24, 1510–1521

H. Shan, J. Kattge, P. B. Reich, A. Banerjee, F. Schrodt, and M. Reichstein. (2012) Gap Filling in the Plant Kingdom – Trait Prediction Using Hierarchical Probabilistic Matrix Factorization . International Conference on Machine Learning (ICML)

See Also

GapFilling, CalculateCvRmse

Examples

## Not run: 
    library(BHPMF)
    data(trait.info)  # Read the matrix X
    data(hierarchy.info) # Read the hierarchy information
	
    # Tune the parameter with default values
    out1 <- TuneBhpmf(trait.info, hierarchy.info)
    num.latent.factors <- out1$best.number.latent.features
	
    # Tune the paramter using partial hierarchy: using only the first 2 level of hierarchy
    out2 <- TuneBhpmf(trait.info, hierarchy.info, used.num.hierarchy.levels=2,
                      num.samples=500, burn=20, gaps=3)
    num.latent.factors <- out2$best.number.latent.features	

## End(Not run)

BHPMF documentation built on June 20, 2017, 9:10 a.m.