mbl {resemble}R Documentation

A function for memory-based learning (mbl)

Description

This function is implemented for memory-based learning (a.k.a. instance-based learning or local regression) which is a non-linear lazy learning approach for predicting a given response variable from a set of predictor variables. For each observation in a prediction set, a specific local regression is carried out based on a subset of similar observations (nearest neighbors) selected from a reference set. The local model is then used to predict the response value of the target (prediction) observation. Therefore this function does not yield a global regression model.

Usage

mbl(Xr, Yr, Xu, Yu = NULL, k, k_diss, k_range, spike = NULL,
    method = local_fit_wapls(min_pls_c = 3, max_pls_c = min(dim(Xr), 15)),
    diss_method = "pca", diss_usage = "predictors", gh = TRUE,
    pc_selection = list(method = "opc", value = min(dim(Xr), 40)),
    control = mbl_control(), group = NULL, center = TRUE, scale = FALSE,
    verbose = TRUE, documentation = character(), seed = NULL, ...)

Arguments

Xr

a matrix of predictor variables of the reference data (observations in rows and variables in columns).

Yr

a numeric matrix of one column containing the values of the response variable corresponding to the reference data.

Xu

a matrix of predictor variables of the data to be predicted (observations in rows and variables in columns).

Yu

an optional matrix of one column containing the values of the response variable corresponding to the data to be predicted. Default is NULL.

k

a vector of integers specifying the sequence of k-nearest neighbors to be tested. Either k or k_diss must be specified. This vector will be automatically sorted into ascending order. If non-integer numbers are passed, they will be coerced to the next upper integers.

k_diss

a numeric vector specifying the sequence of dissimilarity thresholds to be tested for the selection of the nearest neighbors found in Xr around each observation in Xu. These thresholds depend on the corresponding dissimilarity measure specified in the object passed to control. Either k or k_diss must be specified.

k_range

an integer vector of length 2 which specifies the minimum (first value) and the maximum (second value) number of neighbors to be retained when the k_diss is given.

spike

an integer vector (with positive and/or negative values) indicating the indices of observations in Xr that must be either be forced into or avoided in the neighborhoods of every Xu observation. Default is NULL (i.e. no observations are forced or avoided). Note that this argument is not intended for increasing or reducing the neighborhood size which is only controlled by k or k_diss and k_range. By forcing observations into the neighborhood, some of the farthest observations may be forced out of the neighborhood. In contrast, by avoiding observations in the neighborhood, some of farthest observations may be included into the neighborhood. See details.

method

an object of class local_fit which indicates the type of regression to conduct at each local segment as well as additional parameters affecting this regression. See local_fit function.

diss_method

a character string indicating the spectral dissimilarity metric to be used in the selection of the nearest neighbors of each observation. Options are:

  • "pca" (Default): Mahalanobis distance computed on the matrix of scores of a Principal Component (PC) projection of Xr and Xu. PC projection is done using the singular value decomposition (SVD) algorithm. See ortho_diss function.

  • "pca.nipals": Mahalanobis distance computed on the matrix of scores of a Principal Component (PC) projection of Xr and Xu. PC projection is done using the non-linear iterative partial least squares (nipals) algorithm. See ortho_diss function.

  • "pls": Mahalanobis distance computed on the matrix of scores of a partial least squares projection of Xr and Xu. In this case, Yr is always required. See ortho_diss function.

  • "cor": correlation coefficient between observations. See cor_diss function.

  • "euclid": Euclidean distance between observations. See f_diss function.

  • "cosine": Cosine distance between observations. See f_diss function.

  • "sid": spectral information divergence between observations. See sid function.

Alternatively, a matrix of dissimilarities can also be passed to this argument. This matrix is supposed to be a user-defined matrix representing the dissimilarities between observations in Xr and Xu. When diss_usage = "predictors", this matrix must be squared (derived from a matrix of the form rbind(Xr, Xu)) for which the diagonal values are zeros (since the dissimilarity between an object and itself must be 0). On the other hand, if diss_usage is set to either "weights" or "none", it must be a matrix representing the dissimilarity of each observation in Xu to each observation in Xr. The number of columns of the input matrix must be equal to the number of rows in Xu and the number of rows equal to the number of rows in Xr.

diss_usage

a character string specifying how the dissimilarity information shall be used. The possible options are: "predictors", "weights" and "none" (see details below). Default is "predictors".

gh

a logical indicating if the global Mahalanobis distance (in the pls score space) between each observation and the pls mean (centre) must be computed. This metric is known as the GH distance in the literature. Note that this computation is based on the number of pls components determined by using the pc_selection argument. See details.

pc_selection

a list of length 2 used for the computation of GH (if gh = TRUE) as well as in the computation of the dissimilarity methods based on ortho_diss (i.e. when diss_method is one of: "pca", "pca.nipals" or "pls") or when gh = TRUE. This argument is used for optimizing the number of components (principal components or pls factors) to be retained for dissimilarity/distance computation purposes only (i.e not for regression). This list must contain two elements in the following order: method (a character indicating the method for selecting the number of components) and value (a numerical value that complements the selected method). The methods available are:

  • "opc": optimized principal component selection based on Ramirez-Lopez et al. (2013a, 2013b). The optimal number of components (of set of observations) is the one for which its distance matrix minimizes the differences between the Yr value of each observation and the Yr value of its closest observation. In this case value must be a value (larger than 0 and below the minimum dimension of Xr or Xr and Xu combined) indicating the maximum number of principal components to be tested. See the ortho_projection function for more details.

  • "cumvar": selection of the principal components based on a given cumulative amount of explained variance. In this case, value must be a value (larger than 0 and below or equal to 1) indicating the minimum amount of cumulative variance that the combination of retained components should explain.

  • "var": selection of the principal components based on a given amount of explained variance. In this case, value must be a value (larger than 0 and below or equal to 1) indicating the minimum amount of variance that a single component should explain in order to be retained.

  • "manual": for manually specifying a fix number of principal components. In this case, value must be a value (larger than 0 and below the minimum dimension of Xr or Xr and Xu combined). indicating the minimum amount of variance that a component should explain in order to be retained.

The list list(method = "opc", value = min(dim(Xr), 40)) is the default. Optionally, the pc_selection argument admits "opc" or "cumvar" or "var" or "manual" as a single character string. In such a case the default "value" when either "opc" or "manual" are used is 40. When "cumvar" is used the default "value" is set to 0.99 and when "var" is used, the default "value" is set to 0.01.

control

a list created with the mbl_control function which contains additional parameters that control some few aspects of the mbl function (cross-validation, parameter tuning, etc). The default list is as returned by mbl_control(). See the mbl_control function for more details.

group

an optional factor (or character vector vector that can be coerced to factor by as.factor) that assigns a group/class label to each observation in Xr (e.g. groups can be given by spectra collected from the same batch of measurements, from the same observation, from observations with very similar origin, etc). This is taken into account for internal leave-group-out cross validation for pls tuning (factor optimization) to avoid pseudo-replication. When one observation is selected for cross-validation, all observations of the same group are removed together and assigned to validation. The length of the vector must be equal to the number of observations in the reference/training set (i.e. nrow(Xr)). See details.

center

a logical if the predictor variables must be centred at each local segment (before regression). In addition, if TRUE, Xr and Xu will be centred for dissimilarity computations.

scale

a logical indicating if the predictor variables must be scaled to unit variance at each local segment (before regression). In addition, if TRUE, Xr and Xu will be scaled for dissimilarity computations.

verbose

a logical indicating whether or not to print a progress bar for each observation to be predicted. Default is TRUE. Note: In case parallel processing is used, these progress bars will not be printed.

documentation

an optional character string that can be used to describe anything related to the mbl call (e.g. description of the input data). Default: character(). NOTE: his is an experimental argument.

seed

an integer value containing the random number generator (RNG) state for random number generation. This argument can be used for reproducibility purposes (for random sampling) in the cross-validation results. Default is NULL, i.e. no RNG is applied.

...

further arguments to be passed to the dissimilarity function. See details.

Details

The argument spike can be used to indicate what reference observations in Xr must be kept in the neighborhood of every single Xu observation. If a vector of length \(m\) is passed to this argument, this means that the \(m\) original neighbors with the largest dissimilarities to the target observations will be forced out of the neighborhood. Spiking might be useful in cases where some reference observations are known to be somehow related to the ones in Xu and therefore might be relevant for fitting the local models. See Guerrero et al. (2010) for an example on the benefits of spiking.

The mbl function uses the dissimilarity function to compute the dissimilarities between Xr and Xu. The dissimilarity method to be used is specified in the diss_method argument. Arguments to dissimilarity as well as further arguments to the functions used inside dissimilarity (i.e. ortho_diss cor_diss f_diss sid) can be passed to those functions by using ....

The diss_usage argument is used to specify whether the dissimilarity information must be used within the local regressions and, if so, how. When diss_usage = "predictors" the local (square symmetric) dissimilarity matrix corresponding the selected neighborhood is used as source of additional predictors (i.e the columns of this local matrix are treated as predictor variables). In some cases this results in an improvement of the prediction performance (Ramirez-Lopez et al., 2013a). If diss_usage = "weights", the neighbors of the query point (\(xu_{j}\)) are weighted according to their dissimilarity to \(xu_{j}\) before carrying out each local regression. The following tricubic function (Cleveland and Delvin, 1988; Naes et al., 1990) is used for computing the final weights based on the measured dissimilarities:

\[W_{j} = (1 - v^{3})^{3}\]

where if \({xr_{i} \in }\) neighbors of \(xu_{j}\):

\[v_{j}(xu_{j}) = d(xr_{i}, xu_{j})\]

otherwise:

\[v_{j}(xu_{j}) = 0\]

In the above formulas \(d(xr_{i}, xu_{j})\) represents the dissimilarity between the query point and each object in \(Xr\). When diss_usage = "none" is chosen the dissimilarity information is not used.

The global Mahalanobis distance (a.k.a GH) is computed based on the scores of a pls projection. A pls projection model is built with for {Yr}, {Xr} and this model is used to obtain the pls scores of the Xu observations. The Mahalanobis distance between each Xu observation in (the pls space) and the centre of Xr is then computed. The number of pls components is optimized based on the parameters passed to the pc_selection argument. In addition, the mbl function also reports the GH distance for the observations in Xr.

Some aspects of the mbl process, such as the type of internal validation, parameter tuning, what extra objects to return, permission for parallel execution, prediction limits, etc, can be specified by using the mbl_control function.

By using the group argument one can specify groups of observations that have something in common (e.g. observations with very similar origin). The purpose of group is to avoid biased cross-validation results due to pseudo-replication. This argument allows to select calibration points that are independent from the validation ones. In this regard, when validation_type = "local_cv" (used in mbl_control function), then the p argument refers to the percentage of groups of observations (rather than single observations) to be retained in each sampling iteration at each local segment.

Value

a list of class mbl with the following components (sorted either by k or k_diss):

When the k_diss argument is used, the printed results show a table with a column named 'p_bounded. It represents the percentage of observations for which the neighbors selected by the given dissimilarity threshold were outside the boundaries specified in the k_range argument.

Author(s)

Leonardo Ramirez-Lopez and Antoine Stevens

References

Cleveland, W. S., and Devlin, S. J. 1988. Locally weighted regression: an approach to regression analysis by local fitting. Journal of the American Statistical Association, 83, 596-610.

Guerrero, C., Zornoza, R., Gómez, I., Mataix-Beneyto, J. 2010. Spiking of NIR regional models using observations from target sites: Effect of model size on prediction accuracy. Geoderma, 158(1-2), 66-77.

Naes, T., Isaksson, T., Kowalski, B. 1990. Locally weighted regression and scatter correction for near-infrared reflectance data. Analytical Chemistry 62, 664-673.

Ramirez-Lopez, L., Behrens, T., Schmidt, K., Stevens, A., Dematte, J.A.M., Scholten, T. 2013a. The spectrum-based learner: A new local approach for modeling soil vis-NIR spectra of complex data sets. Geoderma 195-196, 268-279.

Ramirez-Lopez, L., Behrens, T., Schmidt, K., Viscarra Rossel, R., Dematte, J. A. M., Scholten, T. 2013b. Distance and similarity-search metrics for use with soil vis-NIR spectra. Geoderma 199, 43-53.

Rasmussen, C.E., Williams, C.K. Gaussian Processes for Machine Learning. Massachusetts Institute of Technology: MIT-Press, 2006.

Shenk, J., Westerhaus, M., and Berzaghi, P. 1997. Investigation of a LOCAL calibration procedure for near infrared instruments. Journal of Near Infrared Spectroscopy, 5, 223-232.

See Also

mbl_control, f_diss, cor_diss, sid, ortho_diss, search_neighbors, local_fit

Examples


library(prospectr)
data(NIRsoil)

# Proprocess the data using detrend plus first derivative with Savitzky and
# Golay smoothing filter
sg_det <- savitzkyGolay(
  detrend(NIRsoil$spc,
    wav = as.numeric(colnames(NIRsoil$spc))
  ),
  m = 1,
  p = 1,
  w = 7
)

NIRsoil$spc_pr <- sg_det

# split into training and testing sets
test_x <- NIRsoil$spc_pr[NIRsoil$train == 0 & !is.na(NIRsoil$CEC), ]
test_y <- NIRsoil$CEC[NIRsoil$train == 0 & !is.na(NIRsoil$CEC)]

train_y <- NIRsoil$CEC[NIRsoil$train == 1 & !is.na(NIRsoil$CEC)]
train_x <- NIRsoil$spc_pr[NIRsoil$train == 1 & !is.na(NIRsoil$CEC), ]

# Example 1
# A mbl implemented in Ramirez-Lopez et al. (2013,
# the spectrum-based learner)
# Example 1.1
# An exmaple where Yu is supposed to be unknown, but the Xu
# (spectral variables) are known
my_control <- mbl_control(validation_type = "NNv")

## The neighborhood sizes to test
ks <- seq(40, 140, by = 20)

sbl <- mbl(
  Xr = train_x,
  Yr = train_y,
  Xu = test_x,
  k = ks,
  method = local_fit_gpr(),
  control = my_control,
  scale = TRUE
)
sbl
plot(sbl)
get_predictions(sbl)

# Example 1.2
# If Yu is actually known...
sbl_2 <- mbl(
  Xr = train_x,
  Yr = train_y,
  Xu = test_x,
  Yu = test_y,
  k = ks,
  method = local_fit_gpr(),
  control = my_control
)
sbl_2
plot(sbl_2)

# Example 2
# the LOCAL algorithm (Shenk et al., 1997)
local_algorithm <- mbl(
  Xr = train_x,
  Yr = train_y,
  Xu = test_x,
  Yu = test_y,
  k = ks,
  method = local_fit_wapls(min_pls_c = 3, max_pls_c = 15),
  diss_method = "cor",
  diss_usage = "none",
  control = my_control
)
local_algorithm
plot(local_algorithm)

# Example 3
# A variation of the LOCAL algorithm (using the optimized pc
# dissmilarity matrix) and dissimilarity matrix as source of
# additional preditors
local_algorithm_2 <- mbl(
  Xr = train_x,
  Yr = train_y,
  Xu = test_x,
  Yu = test_y,
  k = ks,
  method = local_fit_wapls(min_pls_c = 3, max_pls_c = 15),
  diss_method = "pca",
  diss_usage = "predictors",
  control = my_control
)
local_algorithm_2
plot(local_algorithm_2)

# Example 4
# Running the mbl function in parallel with example 2

n_cores <- 2

if (parallel::detectCores() < 2) {
  n_cores <- 1
}

# Alternatively:
# n_cores <- parallel::detectCores() - 1
# if (n_cores == 0) {
#  n_cores <- 1
# }

library(doParallel)
clust <- makeCluster(n_cores)
registerDoParallel(clust)

# Alernatively:
# library(doSNOW)
# clust <- makeCluster(n_cores, type = "SOCK")
# registerDoSNOW(clust)
# getDoParWorkers()

local_algorithm_par <- mbl(
  Xr = train_x,
  Yr = train_y,
  Xu = test_x,
  Yu = test_y,
  k = ks,
  method = local_fit_wapls(min_pls_c = 3, max_pls_c = 15),
  diss_method = "cor",
  diss_usage = "none",
  control = my_control
)
local_algorithm_par

registerDoSEQ()
try(stopCluster(clust))

# Example 5
# Using local pls distances
with_local_diss <- mbl(
  Xr = train_x,
  Yr = train_y,
  Xu = test_x,
  Yu = test_y,
  k = ks,
  method = local_fit_wapls(min_pls_c = 3, max_pls_c = 15),
  diss_method = "pls",
  diss_usage = "predictors",
  control = my_control,
  .local = TRUE,
  pre_k = 150,
)
with_local_diss
plot(with_local_diss)


[Package resemble version 2.2.3 Index]