HEMDAG (version 2.1.3)

TPR-DAG-holdout: TPR-DAG holdout experiments

Description

High level function to correct the computed scores in a hierarchy according to the chosen ensemble algorithm

Usage

Do.TPR.DAG.holdout(threshold = seq(from = 0.1, to = 0.9, by = 0.1),
  weight = seq(from = 0.1, to = 1, by = 0.1), kk = 5, folds = 5,
  seed = 23, norm = TRUE, norm.type = NULL, positive = "children",
  bottomup = "threshold.free", rec.levels = seq(from = 0.1, to = 1, by =
  0.1), n.round = 3, f.criterion = "F", metric = NULL,
  flat.file = flat.file, ann.file = ann.file, dag.file = dag.file,
  ind.test.set = ind.test.set, ind.dir = ind.dir, flat.dir = flat.dir,
  ann.dir = ann.dir, dag.dir = dag.dir, hierScore.dir = hierScore.dir,
  perf.dir = perf.dir)

Arguments

threshold

range of threshold values to be tested in order to find the best threshold (def: from:0.1, to:0.9, by:0.1). The denser the range is, the higher the probability to find the best threshold is, but obviously the execution time will be higher. Set the parameter threshold only for the variants that requiring a threshold for the positive nodes selection, otherwise set the parameter threshold to zero

weight

range of weight values to be tested in order to find the best weight (def: from:0.1, to:0.9, by:0.1). The denser the range is, the higher the probability to find the best threshold is, but obviously the execution time will be higher. Set the parameter weight only for the weighted variants, otherwise set the parameter weight to zero

kk

number of folds of the cross validation (def: kk=5) on which tuning the parameters threshold and weight and

folds

number of folds of the cross validation on which computing the performance metrics averaged across folds (def. 5). If folds=NULL, the performance metrics are computed one-shot, otherwise the performance metrics are averaged across folds.

seed

initialization seed for the random generator to create folds (def. 23). If NULL folds are generated without seed initialization. The parameter seed controls both the parameter kk and the parameter folds.

norm

boolean value: should the flat scores matrix be normalized?

• TRUE (def.): the flat scores matrix has been already normalized in according to a normalization method;

• FALSE: the flat scores matrix has not been normalized yet. See the parameter norm.type to set the on the fly normalization method to apply among those possible.

norm.type

can be one of the following three values:

1. NULL (def.): set norm.type to NULL if and only if the parameter norm is set to TRUE;

2. MaxNorm: each score is divided for the maximum of each class;

3. Qnorm: quantile normalization. preprocessCore package is used.

positive

choice of the positive nodes to be considered in the bottom-up strategy. Can be one of the following values:

• children (def.): for each node are considered its positive children;

• descendants: for each node are considered its positive descendants;

bottomup

strategy to enhance the flat predictions by propagating the positive predictions from leaves to root. It can be one of the following values:

• threshold.free (def.): positive nodes are selected on the basis of the threshold.free strategy (def.);

• threshold: positive nodes are selected on the basis of the threshold strategy;

• weighted.threshold.free: positive nodes are selected on the basis of the weighted.threshold.free strategy;

• weighted.threshold: positive nodes are selected on the basis of the weighted.threshold strategy;

• tau: positive nodes are selected on the basis of the tau strategy. NOTE: tau is only a DESCENS variants. If you use tau strategy you must set the parameter positive=descendants;

rec.levels

a vector with the desired recall levels (def: from:0.1, to:0.9, by:0.1) to compute the the Precision at fixed Recall level (PXR)

n.round

number of rounding digits to be applied to the hierarchical scores matrix (def. 3). It is used for choosing the best threshold on the basis of the best F-measure

f.criterion

character. Type of F-measure to be used to select the best F-measure. Two possibilities:

1. F (def.): corresponds to the harmonic mean between the average precision and recall

2. avF: corresponds to the per-example F-score averaged across all the examples

metric

a string character specifying the performance metric on which to maximize the parametric ensemble variant. It can be one of the following values:

1. PRC: the parametric ensemble variant is maximized on the basis of AUPRC (AUPRC);

2. FMAX: the parametric ensemble variant is maximized on the basis of Fmax (Multilabel.F.measure;

3. NULL: on the threshold.free variant none parameter optimization is needed, since the variant is non-parametric. So, if bottomup=threshold.free set metric=NULL (def.).

flat.file

name of the file containing the flat scores matrix to be normalized or already normalized (without rda extension)

ann.file

name of the file containing the the label matrix of the examples (without rda extension)

dag.file

name of the file containing the graph that represents the hierarchy of the classes (without rda extension)

ind.test.set

name of the file containing a vector of integer numbers corresponding to the indices of the elements (rows) of scores matrix to be used in the test set

ind.dir

relative path to folder where ind.test.set is stored

flat.dir

relative path where flat scores matrix is stored

ann.dir

relative path where annotation matrix is stored

dag.dir

relative path where graph is stored

hierScore.dir

relative path where the hierarchical scores matrix must be stored

perf.dir

relative path where the performance measures must be stored

Value

Two rda files stored in the respective output directories:

1. Hierarchical Scores Results: a matrix with examples on rows and classes on columns representing the computed hierarchical scores for each example and for each considered class. It is stored in the hierScore.dir directory.

2. Performance Measures: flat and hierarchical performace results:

   1. AUPRC results computed though AUPRC.single.over.classes (AUPRC);

   2. AUROC results computed through AUROC.single.over.classes (AUROC);

   3. PXR results computed though PXR.at.multiple.recall.levels.over.classes (PXR);

   4. FMM results computed though compute.Fmeasure.multilabel (FMM);

It is stored in the perf.dir directory.

Details

The parametric hierarchical ensemble variants are cross-validated by maximizing in according to the metric chosen in the parameter metric, that is F-measure (Multilabel.F.measure) or AUPRC (AUPRC).

The function checks if the number of classes between the flat scores matrix and the annotations matrix mismatched. If so, the number of terms of the annotations matrix is shrunk to the number of terms of the flat scores matrix and the corresponding subgraph is computed as well. N.B.: it is supposed that all the nodes of the subgraph are accessible from the root.

See Also

TPR-DAG-variants

Examples

# NOT RUN {
data(graph);
data(scores);
data(labels);
data(test.index);
if (!dir.exists("data")){
	dir.create("data");
}
if (!dir.exists("results")){
	dir.create("results");
}
save(g,file="data/graph.rda");
save(L,file="data/labels.rda");
save(S,file="data/scores.rda");
save(test.index, file="data/test.index.rda");
ind.dir <- dag.dir <- flat.dir <- ann.dir <- "data/";
hierScore.dir <- perf.dir <- "results/";
dag.dir <- flat.dir <- ann.dir <- "data/";
ind.test.set <- "test.index";
dag.file <- "graph";
flat.file <- "scores";
ann.file <- "labels";
threshold <- weight <- 0;
norm.type <- "MaxNorm";
positive <- "children";
bottomup <- "threshold.free";
rec.levels <- seq(from=0.1, to=1, by=0.1);
Do.TPR.DAG.holdout(threshold=threshold, weight=weight, kk=5, folds=NULL, seed=23, norm=FALSE, 
norm.type=norm.type, positive=positive, bottomup=bottomup, rec.levels=rec.levels, n.round=3, 
f.criterion="F", metric=NULL, flat.file=flat.file, ann.file=ann.file, dag.file=dag.file, 
ind.test.set=ind.test.set, ind.dir=ind.dir, flat.dir=flat.dir, ann.dir=ann.dir, 
dag.dir=dag.dir, hierScore.dir=hierScore.dir, perf.dir=perf.dir);
# }