Changes in version 3.4.0
o
Companion website with more in-depth explanation and examples.
o
Default random forest classifier based on ranger now does
two-step classification; one for variable importance and one
for model fitting, as recommended by ranger's developer.
o
More functions use automatic parameter value selection as their
defaults.
o
randomSelection function to choose random sets of features in
cross-validation.
o
crossValidate function now permits custom parameter tuning via
extraParams.
o
Elastic net GLM and ordinary GLM now calculate class weights be
default, so as to perform well in class-imbalanced scenarios.
o
precisionPathwaysTrain and precisionPathwaysPredict functions
for building tree-like models of multiple assays and their
accessory functions calcCostsAndPerformance, bubblePlot,
flowchart, strataPlot for model performance evaluation.
o
crissCrossValidate function that takes a list of data sets with
the same set of features and the same set of outcomes and does
all possible pairs of training and prediction to evaluate
generalisability. crissCrossPlot for visual evaluation.
o
Changes in version 3.2.0
o
Fast Cox survival analysis.
o
Simple parameter sets, as used by crossVaildate, now come with
tuning parameter grid as standard.
o
Wrappers are greatly simplified. Now, there is only one method
for a data frame and they are not exported because they are not
used directly by the end-user anyway.
o
prepareData function to filter and subset input data using
common ways, such as missingness and variability.
o
Invalid column names of data (e.g. spaces, hyphens) are
converted into safe names before modelling but converted back
into original names for tracking ranked and selected features.
o
available function shows the keywords corresponding to
transformation, selection, classifier functions.
o
More functions have automatically-selected parameters based on
input data, reducing required user-specified parameters.
o
New classifiers added for random survival forests and extreme
gradient boosting.
o
Adaptive sampling for modelling with uncertainty of class
labels can be enabled with adaptiveResamplingDelta.
o
Parameter tuning fixed to only use samples from the training
set.
Changes in version 3.0.0
o
Now supports survival models and their evaluation, in addition
to existing classification functionality.
o
Cross-validation no longer requires specific annotations like
data set name and classifier name. Now, the user can specify
any characteristics they want and use these as variables to
group by or change line appearances. Also, characteristics like
feature selection name and classifier name are automatically
filled in from an internal table.
o
Ease of use greatly inproved by crossValidate function which
allows specification of classifiers by a single keyword.
Previously, parameter objects such as SelectParams and
TrainParams had to be explicitly specified, making it
challenging for users not familar with S4 object-oriented
programming.
o
Basic multi-omics data integration functionality available via
crossValidate which allows combination of different tables.
Pre-validation and PCA dimensionality techniques provide a fair
way to compare high-dimensional omics data with low-dimensional
clinical data. Also, it is possible to simply concatenate all
data tables.
o
Model-agnostic variable importance calculated by training when
leaving out one selected variable at a time. Turned off by
default as it substantially increases run time. See
doImportance parameter of ModellingParams for more details.
o
Parameters specifying the cross-validation procedure and data
modelling formalised as CrossValParams and ModellingParams
classes.
o
Feature selection can now be done either based a on
resubstitution metric (i.e. train and test on the training
data) or a cross-validation metric (i.e. split the training
data into training and testing partitions to tune the selected
features). all feature selection functions have been converted
into feature ranking functions, because the selection procedure
is a feature of cross-validation.
o
All function and class documentation coverted from manually
written Rd files to Roxygen format.
o
Human Reference Interactome (binary experimental PPI) included
in bundled data for pairs-based classification. See ?HuRI for
more details.
o
Performance plots can now do either box plots or violin plots.
Box plot remains the default style.
Changes in version 2.14.0
o
Upsampling and downsampling to equalise class sizes added.
Changes in version 2.6.0
o
Two-stage easy-hard classifier added.
Changes in version 2.2.0
o
getClasses is no longer a slot of PredictParams. Every
predictor function needs to return either a factor vector of
classes, a numeric vector of class scores for the second class,
or a data frame with a column for the predicted classes and
another for the second-class scores.
o
Cross-validations which use folds ensure that samples belonging
to each class are in approximately the same proportions as they
are for the entire data set.
o
Classification can reuse fitted model from previous
classification by using previousTrained function.
o
Feature selection using gene sets and networks. Classification
can use meta-features derived from the individual features used
for feature selection.
o
tTestSelection function for feature selection based on ordinary
t-test statistic ranking. Now the default feature selection
function, if none is specified.
o
Tuning parameter optimisation metric is specified by providing
a tuneOptimise parameter to TrainParams rather than depending
on ResubstituteParams being used during feature selection.
Changes in version 2.0.0
o
Broad support for DataFrame and MultiAssayExperiment data sets
by feature selection and classification functions.
o
The majority of processing is now done in the DataFrame method
for functions that implement methods for multiple kinds of
inputs.
o
Elastic net GLM classifier and multinomial logistic regression
classifier wrapper functions.
o
Plotting functions have a new default style using a white
background with black axes.
o
Vignette simplified and uses a new mass cytometry data set with
clearer differences between classes to demonstrate
classification and its performance evaluation.
Changes in version 1.12.0
o
Alterations to make plots compatible with ggplot versions 2.2
and greater.
o
calcPerformance can calculate some performance metrics for
classification tasks based on data sets with more than two
classes.
o
Sample-wise metrics, like sample-specific error rate and
sample-specific accuracy are calculated by calcPerformance and
added to the ClassifyResult object, rather than by
samplesMetricMap and being inaccessible to the end-user.
Changes in version 1.10.0
o
errorMap replaced by samplesMetricMap. The plot can now show
either error rate or accuracy.
Changes in version 1.8.0
o
Ordinary k-fold cross-validation option added.
o
Absolute difference of group medians feature selection function
added.
Changes in version 1.4.0
o
Weighted voting mode that uses the distance from an observation
to the nearest crossover point of the class densities added.
o
Bartlett Test selection function included.
o
New class SelectResult. rankPlot and selectionPlot can
additionally work with lists of SelectResult objects. All
feature selection functions now return a SelectResult object or
a list of them.
o
priorSelection is a new selection function for using features
selected in a prior cross validation for a new data set
classification.
o
New weighted voting mode, where the weight is the distance of
the x value from the nearest crossover point of the two
densities. Useful for predictions with skewed features.
Changes in version 1.2.0
o
More classification flexibility, now with parameter tuning
integrated into the process.
o
New performance evaluation functions, such as a ROC curve and a
performance plot.
o
Some existing predictor functions are able to return class
scores, not just class labels.
Changes in version 1.0.0
o
First release of the package, which allows parallelised and
customised classification, with many convenient performance
evaluation functions.