Several pre-processing steps, i.e. transformation, normalisation, batch normalisation, imputation and clustering have been re-implemented as objects. This allows for better portability between experiments, improved flexibility and extensibility to newer methods, and better forward compatibility.
Variable importance / feature selection is now implemented in an
object-oriented fashion. Variable importance data written to the file
system are now vimpTable objects, instead of data collected
in a loose list. The aggregate_vimp_table and
get_vimp_table methods can be used to aggregate multiple
variable importance tables and retrieve the variable importance table as
a data.table respectively.
It is now possible to warm-start different experiments by
providing an experimentData object to the
summon_familiar function using the
experiment_data argument. Such objects are created using
the precompute_data_assignment,
precompute_feature_info and precompute_vimp
functions. These functions assign instances to training, validation and
test data, derive feature processing parameters and compute variable
importance respectively. precompute_vimp and
precompute_feature_info includes any previous
steps.
It is now possible to set the number of bootstraps that are
initially explored for hyperparameter optimisation using the
smbo_initial_bootstraps configuration parameter. The
default value is 1, indicating that the initial
hyperparameter sets are initially evaluated on a single
bootstrap.
Updated objects now both show initial and current familiar versions.
Consistency of S4 objects is more actively checked in exported
methods and functions by calling update_object. This
prevents errors attempting when attempting to use a more recent version
of familiar than the one used to create the objects. Previously these
checks were only rigorously performed for objects that were loaded
internally, and not those directly loaded by the user.
The n_dim parameter for isolation forests is now
correctly set for datasets without any features.
Pseudo-R2 similarity metrics now correctly produce a value of 1.0 for exact fits. Previously these could produce infinite log-likelihoods, and return default value of 0.0 (no similarity).
A default value of the time_max parameter can now be
determined even if training cohorts are not explicitly set.
Fixed an issue with LASSO-based imputation when aggregating from LASSO models with different required features. This is a temporary solution, that will be tackled more comprehensively in version 1.2.0.
Fixed an issue with the parenthesis check in nested experimental designs.
Fixed an issue that would prevent data where the same instance appears multiple times (e.g. bootstraps) from being properly evaluated.
signature_only and
random feature selection methods._default. This is mainly
done to avoid long hyperparameter optimisation times for such methods
during feature selection.Hyperparameter time limit and time taken are now always correctly parsed to minutes instead of seconds, minutes, or rarely, hours.
Familiar model objects now correctly show that no novelty
detector was trained with novelty_detector="none".
Fixed missing documentation for the
optimisation_function configuration parameter.
Confidence intervals in calibration plots are no longer cropped to the [0,1] range. Previously, the estimate itself was not cropped, whereas its confidence interval was.
xgboost models now still work after being loaded
from a file.
train_familiar function that trains (and returns)
models, but skips evaluation steps. This function is essentially a
wrapper around summon_familiar.multivariate_regression, mrmr and
lasso now respect signature features set using the
signature configuration parameter. Features provided in
signature are always selected for the resulting model, and
were therefore ignored during feature selection for both univariate and
multivariate method. This has changed, so that multivariate methods now
use the signature features as the basic set and attempt to identify any
additional suitable features. Signature features are still ignored for
univariate methods.successive_halving or
stochastic_reject as exploration methods. Fewer bootstraps
are assessed during intensification steps if there are any bootstraps
that have only partially been sampled by the hyperparameter sets under
evaluation. This should accelerate the optimisation process
considerably. The (so far untested) rationale is that the hyperparameter
learners should generally be able to accurately model the optimisation
score of hyperparameter sets using locally sparse data.
smbo_time_limit parameter.
Optimisation will stop once this limit has been exceeded. Note that
familiar does not actively kill ongoing optimisation processes, but
waits until they complete before stopping optimisation. Actively killing
processes would require a general overhaul of the parallelisation
routines used in familiar, which is complex and not an urgent
priority.summary, vcov and coef
method for familiarModel objects. These respectively apply
summary, vcov and coef to the
stored model.optimisation_function to determine optimisation and overall
summary scores of hyperparameter sets. Newly introduced are:
validation_minus_sd: The mean performance on
out-of-bag data minus its standard deviation.
validation_25th_percentile: The 25th percentile of
model performance on out-of-bag data.
model_estimate: The estimated model performance
inferred by the hyperparameter model that was previously used to
identify new candidate hyperparameter sets. Not available for random
search.
model_estimate_minus_sd: The estimated model
performance minus its standard deviation. Not available for random
search.
optimisation_function is now
validation.smbo_stop_tolerance now depends on the
number of samples and varies between 0.01 for 100 or fewer
samples, and 0.001 for 10000 or more samples.Fixed some missing verbosity settings.
Fixed a deprecation warning in the xgboost package for DART boosters, that appeared for versions 1.4.0 and newer.
Trained models now contain information for features used only for missing data inference.
Capture a rare error produced by the maxstat package when determining optimal risk stratification thresholds.
Fixed unexpected behaviour in interactions between project_dir and experiment_dir configuration parameters.
Fixed a bug that prevented computation of model variable importance when data was not loaded. This affected learners based on glmnet, randomForestSRC and ranger packages.
Fixed a bug that prevented familiar from identifying iteration files that were created previously for the same experiment.
Fixed a bug that would only occur during unit testing that was due to feature importance files being considered present without these files actually existing.
All models are now trimmed to remove unnecessary objects such as nested environments, copies of training data, etc. This should cause an overall smaller memory footprint. Note that this does not necessarily imply anonymisation of the data. Notably, k-nearest neighbour learners still maintain a copy of the training dataset internally.
Naive Bayes and k-nearest neighbour learners now use the
e1071 package instead of klaR, which has been
deprecated from familiar. Some hyperparameters changed accordingly. See
the learners vignette.
Added individual conditional expectation and partial dependence plots. These plots show the response of a model across a range of values for a particular feature.
Hyperparameter optimisation now allows for more flexibility in
setting the exploration method. The exploration method determines how
less promising hyperparameter sets are pruned during intensification
steps. The exploration method can be set using the
exploration_method argument. Familiar currently supports
the following options:
successive_halving (default): The set of alternative
parameter sets is pruned by removing the worst performing half of the
sets after each step. The set of investigated parameter sets gets
progressively smaller.
stochastic_reject: The set of alternative parameter
sets is pruned by comparing the performance of each parameter set with
that of the incumbent best parameter set using a paired
Wilcoxon test. This was the previous default.
none: The set of alternative parameter sets is not
pruned.
We can now also change the hyperparameter learner used to infer
suitability of candidate hyperparameter sets for further exploration.
The learner can be set using the hyperparameter_learner
argument, and supports the following options:
gaussian_process (default): Creates a localised
approximate Gaussian deterministic Gaussian Processes.
bayesian_additive_regression_trees or
bart: Uses Bayesian Additive Regression Trees for
inference. Unlike standard random forests, BART allows for estimating
posterior distributions directly and can extrapolate.
random_forest: Creates a random forest for
inference. A weakness of random forests is their lack of extrapolation
beyond observed values, which limits their usefulness in exploiting
promising areas of hyperparameter space. This was the previously
supported option.
random or random_search: Forgoes the
use of models to steer optimisation. Instead, a random search is
performed. This means the hyperparameter space is sampled at
random.
Three new vignettes have been added. The first vignette is an introductory vignette on how to get started with familiar. The second vignette describes how familiar can be used prospectively. The third vignette describes evaluation and explanation steps in familiar and how they are implemented. All other vignettes have been reviewed and updated.
A predict method was added to allow for direct
inference of estimated values for one or more instances. Functionality
is the same as other predict methods, but data
(newdata) should always be provided. Familiar does not
store development data with the model.
The RServe backend was retired, as recent versions of the RServe and RSclient packages produced errors that could not be resolved.
A sample_limit parameter was added to limit the
number of samples used during evaluation. This parameter can be
specified for the sample_similarity evaluation
step.
Class levels for categorical outcomes, if not explicitly
specified using the class_levels parameter, are now sorted
before being set based on the data. Previously class levels would be set
based on order of appearance.
Added mean, mean_trim and
mean_winsor as methods to set risk-group stratification
thresholds.
Novelty detector algorithms are now implemented as S4 classes. This was primarily done to make it easier to add additional methods.
novelty_detector and
detector_parameters configuration parameters were
added.
Isolation forests are now grown with a decreased memory footprint.
Presence of instances with a survival time of 0 or lower will produce a warning. Though familiar itself will handle such instances just fine, other packages will produce errors. Since familiar captures such errors and handles them internally, a warning is provided to indicate potential issues.
The hyperparameter optimisation algorithm performs an improved search of the local neighbourhood of good hyperparameter sets. Previously large parts of local neighbourhoods were ignored as their utility may not have exceeded that of the seed set. The algorithm is now no longer myopic. Instead, the seed set is used as starting point for exploration, and a random path through hyperparameter space is charted. The most promising hyperparameter sets are chosen after repeating this procedure several times and for several seed points.
The smbo_intensify_stop_p_value parameter was
renamed to smbo_stochastic_reject_p_value.
The eval_times argument used in plot and export
methods was renamed to evaluation_times to match that of
the synonymous configuration parameter.
Receiver operating characteristic and precision-recall curves are now plotted exactly in circumstances that allow for it. Previously an interpolated version of the curves was always shown.
The default divergent palettes used for feature cluster and sample cluster heatmaps now diverge to white instead of black.
Familiar will now actively check whether packages are installed.
The default number of cores used for parallel processing is now 2
by default, instead of all cores - 1. This can be changed by setting the
parallel_nr_cores parameter.
detail_level, sample_limit,
estimation_type and aggregate_results
arguments used during evaluation and explanation are now propagated from
familiarModel and familiarEnsemble objects
instead of reverting to default values.
Fixed a bug that caused features not incorporated in models to not be exported for the purpose of reporting (aggregate) variable importance.
Fixed a bug that caused clustered features not to be exported for the purpose of reporting (aggregate) variable importance.
Fixed a bug that caused model-based variable importance to be
exported when calling the export_fs_vimp method with any
object that is not a familiarCollection
object.
Fixed a bug that prevented rank aggregation method and thresholds
from being set while exporting variable importance using
export_fs_vimp or export_model_vimp.
Fixed an error that occurred when attempting to fit calibration
data with NA values.
Fixed an error that occurred when attempting to interpolate survival probabilities when collecting pooled data.
Fixed an issue with models from the glmnet package
not training for rare classes / events or censoring in small
datasets.
Fixed an issue that could cause main panels in composite plots that consist of one row of facets and have a legend guide to be have the same height as the guide.
Fixed two issues that could cause errors when plotting a single survival curve.
Points on receiver operating characteristic and precision-recall curves are now always correctly ordered.
Fixed a bug that would prevent hpo_metric,
vimp_aggregation_method and
vimp_aggregation_rank_threshold arguments from being set as
a function argument.
Fixed an error when trying to pass undeclared arguments to the
..train method of familiarModel objects using
callNextMethod().
Added novelty detection. An isolation forest is created at the same time as the main model using the same data. It can then be used to (prospectively) identify samples that are dissimilar to the training samples, and for which the model may need to extrapolate.
novelty_features parameter that can be used to
specify features that should be used for novelty detection, in addition
to those already used in the model.Added update_object methods that allow for backward
compatibility when updating slots of respective objects.
Added support for series-like data. These can be time series, or
multiple measurements were the outcome of interest may change.
Subsampling, e.g. through cross-validation or bootstraps still respects
samples. This means that different series instances of the same sample
are always kept together for subsampling. The series column in the data
set can be set using the series_id_column parameter. This
required changes to what iteration data is stored. This should not cause
any issues with post-hoc analyses, but is not not backward
compatible when updating familiar prior to completing the
modelling and evaluation process.
Improved flexibility of the evaluation process that is conducted to explain and assess models:
Added dynamic_model_loading parameter that supports
dynamic loading of models to an ensemble. This reduces the memory
footprint at the cost of IO overhead as the models are read from the
disk or network when required. By default, all models are statically
attached to an ensemble.
Added skip_evaluation_elements parameter that allows
skipping one or more steps of the evaluation process. This is useful if
some evaluations are not relevant.
Updated parallel_hyperparameter_optimisation and
parallel_evaluation to allow for specifying whether
parallelisation should take place inside (inner) or outside
(outer) the respective processes. For outer
the parallelisation takes place over different subsamples, learners,
etc. This may provide an increase in processing speed, at the cost of
less feedback and a higher memory footprint.
All evaluation steps now produce familiarDataElement
elements. This is not not backward compatible.
familiarData and familiarCollection objects
created using previous versions need to be created anew. enables
flexibility in terms of how data and models are handled for computation.
Specifically, many evaluation steps that focus on models can be
evaluated at the ensemble (ensemble), model
(model) or an intermediate level (hybrid) by
setting the detail_level parameter. The hybrid
level differs from ensemble in that the individual model
predictions are used instead of the prediction of the ensemble itself.
Likewise, the type of estimation can now be flexibly chosen for several
evaluation steps by setting the estimation_type
parameter.
Several parameters are now deprecated:
compute_model_data has been completely deprecated.
This can now be specified using the new detail_level
parameter.
compute_model_ci has been completely deprecated.
This can now be specified using the new estimation_type
parameter.
compute_ensemble_ci has been completely deprecated.
This can now be specified using the new estimation_type
parameter.
aggregate_ci has been replaced by
aggregate_results. Aside from bootstrap confidence
intervals, underlying results for bias-corrected estimates can now be
aggregated.
Calibration plots have been completely revamped to now include
confidence intervals. Moreover, calibration plots based on
bootstrap_confidence_interval and
bias_corrected estimation types no longer show points, but
are based on interpolation to a regular grid after computing a loess
model. point estimates are unaltered. Density plots have
also been revised to use a fixed standard deviation of 0.075, which
prevents some of the erratic behaviour seen previously when most
expected probability values were clustered closely.
Clustering based on feature similarity during the evaluation
process can now be specified after similarity has been computed,
i.e. through export_feature_similarity and
plot_feature_similarity. This allows for changing
clustering parameters after the analysis.
Parallel processing now supports mini-batching. Many processes are actually fast to compute and repeated IO to cluster nodes noticably slows down the process. Mini-batches transfers data to nodes in one go for local sequential processing. In addition, processes that can mini-batch are now measured, and an optimal number of nodes is selected based on IO and process times. This should significantly speed up processes with low process time compared to IO time.
Hyperparameter optimisation now predicts the time taken for training using a specific set of hyperparameters. It uses the predicted time to optimise assignment to nodes for parallel processing. This eliminates an issue where hyperparameter optimisation with parallel nodes could take significantly longer than simple sequential optimisation. In addition, hyperparameter optimisation now has several new or changed configuration parameters:
smbo_random_initialisation is no longer a logical
(TRUE or FALSE) but takes
fixed_subsample, fixed or random
as values. fixed_subsample generates initial
hyperparameters from the same default hyperparameter grid as
fixed, but unlike fixed does not exhaustively
search all options. random creates random sets of initial
hyperparameters.
smbo_n_random_sets can now be used to set the number
of hyperparameters sets created for the fixed_subsample and
random methods.
kernlab is no longer used as backend for computing
support vector machine learners due to lack of stability. Unit testing
showed consistent freezing. We now use the SVM of e1071
which we found to be stable. SVM models created using previous versions
of familiar are no longer compatible.
The default method for bootstrap confidence intervals
(bootstrap_ci_method) is now the percentile
(percentile) method, which replaces the bias-corrected
(bc) method.
The value returned for the bias-corrected bootstrap confidence interval method is now the bias-corrected median, not the point estimate. This harmonises the behaviour of the percentile and bias-corrected confidence interval methods. The bias-corrected median can be viewed as an optimism correction of the point estimate.
Several attribute slots for S4 familiarModel,
familiarEnsemble, familiarData and
familiarCollection objects were removed, revised or added.
Changes are backward compatible due to the new
update_object method.
Lambda parameters of Box-Cox and Yeo-Johnson transformations are
now determined using stats::optimise. The previous, fixed,
settings were sensible for Box-Cox, but the Yeo-Johnson method benefits
from a wider selection. This does not affect backward
compatibility.
The as_data_object method can now be used with
familiarModel and familiarEnsemble objects to
check whether the input data can be correctly formatted, and will
provide meaningful errors if not.
Added show methods for objects that are typically
written to drive, such as familiarModel,
familiarEnsemble, familiarData and
familiarCollection objects.
Added plot_auc_precision_recall_curve method to plot
precision-recall curves.
Random forests created using the rfsrc package are
now anonymous forests, i.e. training data are not stored with the model.
In addition, we explicitly generate and store a random seed, so that the
forest can be regrown for determining variable importance.
Fixed an error that would cause hyperparameter optimisation to not select the optimal set of hyperparameters.
Fixed an error that would cause feature selection to fail when all features in the data are also set to be in the signature.
Fixed an error that occurred when attempting to create risk groups from models that were not successfully trained.
Fixed an error in ComBat batch normalisation caused by invariant or NA features in one or more batches.
Fixed a bug that would incorrectly assign samples to wrong subsamples (e.g. in-bag or out-of-bag data). This only occurred if the same sample identifier exists in different batches.
Fixed an error that occurred prior to hyperparameter optimisation because a model-dependent hyperparameter required to create a metric object may not have been set.
Fixed an error when attempting to perform parallel processing with familiar installed on a non-standard library path.
Fixed an issue where the verbose argument was not
respected when forming clusters of features.
Fixed an issue where absence of censoring for time-to-event data would lead to models not being created.
Fixed an issue where NA would not be removed from the results in
the extract_from_slot function.
Fixed an issue where information would be missing in
familiarEnsemble objects because the first
familiarModel in the ensemble was not trained. This
information cannot be added retroactively.
Fixed an issue that would cause
export_permutation_vimp to export the wrong data when
called by the user.
Fixed an issue that would cause an error when a decision curve was plotted with a confidence interval but without requiring multiple line colours.
Fixed an issue that would cause an error when a receiver operating characteristic curve was plotted with a confidence interval but without requiring multiple line colours.
Fixed an error that occurred when categorical outcome levels are
numeric, e.g. 0 and 1. For some learners, such as svm, this
caused an indexing error as the outcome levels were interpreted as
indices instead of column names.
Fixed an error during batch normalisation when there is a mismatch between features with feature information and actual features present in the dataset.
Fixed an issue that would cause an error when computing a pseudo-R2 similarity score between two arrays that each have one unique value.
All metrics are now implemented as S4 objects, with associated methods. Moreover all metrics now have unit tests.
All plotting algorithms now have unit tests which should increase
stability of the code. Resulting code fixes are not backward
compatible: you may need to recreate the
familiarData objects for binomial endpoints.
Hyperparameter optimisation now has additional parameters:
optimisation_determine_vimp: Allows for determining
variable importance for each of the bootstraps used during
hyperparameter optimisation to avoid positive biases.optimisation_function: replaces the objective
parameter.smbo_stop_tolerance: tolerance for a optimisation score
to be convergent.acquisition_function: an acquisition function can now
be selected.Hyperparameter optimisation can now be performed using multiple optimisation metrics instead of one.
Data computation for individual models can now be explicitly set
using the compute_model_data parameter.
Many bugs were fixed.