Feature selection methods in familiar measure variable importance in a univariate or multivariate setting.

Overview of feature selection methods. ^a This is a general method where an appropriate specific method will be chosen, or multiple distributions or linking families are tested in an attempt to find the best option. ^b This method requires hyperparameter optimisation.
method	tag	binomial	multinomial	continuous	count	survival
correlation
Pearson’s r	`pearson`			×	×	×
Spearman’s ρ	`spearman`			×	×	×
Kendall’s τ	`kendall`			×	×	×
concordance
concordance^a	`concordance`	×	×	×	×	×
CORElearn
information gain ratio	`gain_ratio`	×	×
gini-index	`gini`	×	×
minimum description length	`mdl`	×	×
ReliefF with exponential weighting of distance ranks	`relieff_exp_rank`	×	×	×	×
mutual information
mutual information maximisation	`mim`	×	×	×	×	×
mutual information features selection	`mifs`	×	×	×	×	×
minimum redundancy maximum relevance	`mrmr`	×	×	×	×	×
univariate regression
univariate regression	`univariate_regression`	×	×	×	×	×
multivariate regression
multivariate regression	`multivariate_regression`	×	×	×	×	×
lasso regression
general^a	`lasso`	×	×	×	×	×
logistic	`lasso_binomial`	×
multi-logistic	`lasso_multinomial`		×
normal (gaussian)	`lasso_gaussian`			×
poisson	`lasso_poisson`				×
cox	`lasso_cox`					×
ridge regression
general^a	`ridge`	×	×	×	×	×
logistic	`ridge_binomial`	×
multi-logistic	`ridge_multinomial`		×
normal (gaussian)	`ridge_gaussian`			×
poisson	`ridge_poisson`				×
cox	`ridge_cox`					×
elastic net regression
general^a,b	`elastic_net`	×	×	×	×	×
logistic^b	`elastic_net_binomial`	×
multi-logistic^b	`elastic_net_multinomial`		×
normal (gaussian)^b	`elastic_net_gaussian`			×
poisson^b	`elastic_net_poisson`				×
cox^b	`elastic_net_cox`					×
random forest (RFSRC) variable importance
permutation^b	`random_forest_permutation`	×	×	×	×	×
permutation (unoptimised)	`random_forest_permutation_default`	×	×	×	×	×
minimum depth^b	`random_forest_minimum_depth`	×	×	×	×	×
minimum depth (unoptimised)	`random_forest_minimum_depth_default`	×	×	×	×	×
variable hunting^b	`random_forest_variable_hunting`	×	×	×	×	×
variable hunting (unoptimised)	`random_forest_variable_hunting_default`	×	×	×	×	×
hold-out ^b	`random_forest_holdout`	×	×	×	×	×
hold-out (unoptimised)	`random_forest_holdout_default`	×	×	×	×	×
random forest (ranger) variable importance
permutation^b	`random_forest_ranger_permutation`	×	×	×	×	×
permutation (unoptimised)	`random_forest_ranger_permutation_default`	×	×	×	×	×
hold-out permutation^b	`random_forest_ranger_holdout_permutation`	×	×	×	×	×
hold-out perm. (unoptimised)	`random_forest_ranger_holdout_permutation_default`	×	×	×	×	×
impurity^b	`random_forest_ranger_impurity`	×	×	×	×	x
impurity (unoptimised)	`random_forest_ranger_impurity_default`	×	×	×	×	x
special methods
no selection	`none`	×	×	×	×	×
random selection	`random`	×	×	×	×	×
signature only	`signature_only`	×	×	×	×	×

Configuration options

Feature selection methods and related options can be provided within the feature_selection tag in the xml file or as function argument.

tag / argument	description	default
`fs_method`	The desired feature selection method. Multiple selection methods may be provided at the same time. This setting has no default and must be provided.	– (required)
`fs_method_parameter`	Several feature selection methods have hyperparameters that can be set and/or optimised.	– (optional)
`vimp_aggregation_method`	The aggregation method used to aggregate feature ranks over different bootstraps.	`borda`
`vimp_aggregation_rank_threshold`	Several aggregation methods count features if they have a rank below the threshold, i.e. are among the most important features. If `NULL`, a dynamic threshold is decided through Otsu-thresholding.	`NULL`
`parallel_feature_selection`	Enables parallel processing for feature selection. Ignored if `parallel=FALSE`.	`TRUE`

Providing parameters for feature selection

Some of the feature selection methods, notably those based on random forests and (penalised) regression, have parameters that can be set. These parameters are mentioned under the respective entries in the Overview of feature selection methods section. Moreover, some of these parameters are model parameters. In this case, these parameters are optimised using hyperparameter optimisation, which is described in the learning algorithms and hyperparameter optimisation vignette.

The syntax for such parameters is the same as for hyperparameter optimisation. For the multivariate_regression feature selection method the alpha parameter (which determines feature drop-out during forward selection) may be provided as follows using the configuration file:

<fs_method_parameter>
  <multivariate_regression>
    <alpha>0.05</alpha>
  </multivariate_regression>
</fs_method_parameter>

Or as a nested list passed as the fs_method_parameter argument to summon_familiar:

fs_method_parameter = list("multivariate_regression"=list("alpha"=0.05))

Overview of feature selection methods

The feature selection methods implemented in familiar are described in more detail in this section.

Correlation methods

Correlation methods determine variable importance by assessing the correlation between a feature and the outcome of interest. High (anti-)correlation indicates an important feature, whereas low (anti-)correlation indicates that a feature is not directly related to the outcome. Correlation-based variable importance is determined using the cor function of the stats package that is part of the R core distribution (R Core Team 2019).

Three correlation coefficients can be computed:

pearson: Pearson’s \(r\)
spearman: Spearman’s \(\rho\)
kendall: Kendall’s \(\tau\)

To compute correlation of features with survival outcomes, only samples with an event are considered.

Concordance methods

Concordance methods assess how well the ordering of feature values corresponds to the ordering of the outcome. The method internally refers to the gini method for binomial and multinomial outcomes and to the kendall method for continuous and count outcomes. For survival outcomes, concordance is measured using the concordance_index.

CORElearn methods

Familiar provides an interface to several feature selection methods implemented in the CORElearn package. These methods are the Information Gain Ratio (gain_ratio), the Gini-index (gini), Minimum Description Length (mdl) and ReliefF and rReliefF with exponential distance rank weighting (relieff_exp_rank).

Mutual information-based methods

Mutual information \(I\) is a measure of interdependency between two variables \(x\) and \(y\). In the context of feature selection, \(x\) is a feature vector and \(y\) is the outcome vector.

Computing mutual information requires that probability distributions of \(x\) and \(y\) are known. In practice we don’t know either one. For categorical \(x\) and \(y\) we can use the sample estimates instead. For continuous or mixed data, the situation is more complex.

In familiar we therefore use the following three approaches to compute mutual information:

For binomial and multinomial outcomes mutual information is computed using sample estimates. In case of continuous \(x\), these are discretised into \(\lceil 2 n^{1/3} \rceil\) bins, with \(n\) the number of samples, after which computation is conducted as if \(x\) was a categorical variable.
For continuous and count outcomes, we use the approximation proposed by De Jay et al. after Gel’fand and Yaglom (Gel′fand and Yaglom 1959; De Jay et al. 2013): \(I = -0.5 \log(1 - \rho(x,y)^2 + \epsilon)\), with \(\rho(x,y)\) Spearman’s correlation coefficient and \(\epsilon\) a small positive number to prevent \(\log(0)\).
For survival outcomes the second method is adapted for use with a concordance index: \(I = -0.5 \log(1 - (2 * (ci-0.5))^2 + \epsilon)\), with \(ci\) the concordance index.

We opted to adapt the approach based on the outcome type as this ensures that a single consistent approach is used to assess all feature data in an analysis, thus making results comparable.

Mutual information maximisation

The mim method is a univariate method that ranks each feature by its mutual information with the outcome.

Mutual information feature selection

Mutual information feature selection (MIFS) finds a feature set that maximises mutual information (Battiti 1994). This is done using forward selection. As in mutual information maximisation, mutual information \(I_{y,j}\) between each feature and the outcome is computed. Starting from a potential pool of all features, the feature with the highest mutual information is selected and removed from the pool.

The rest proceeds iteratively. The mutual information \(I_{s,1j}\) between the previously selected feature and the remaining features is computed. This mutual information is also called redundancy. The feature with the highest mutual information with the outcome and least redundancy (i.e. maximum \(I_{y,j} - I_{s,1j}\)) is selected next, and removed from the pool of remaining features. Then the mutual information \(I_{s,2j}\) between this feature and remaining features is computed, and the feature that maximises \(I_{y,j} - I{s,1j} - I_{s,2j}\) is selected, and so forth.

The iterative process stops if there is no feature \(j\) for which \(I_{y,j} - \sum_{i\in S} I_{s,ij} > 0\), with \(S\) being the subset of selected features, or all features have been exhausted.

To reduce the number of required computations, the implementation in familiar actively filters out any feature \(j\) for which \(I_{y,j} - \sum_{i\in S} I_{s,ij} \leq 0\) at the earliest instance, as the \(\sum_{i\in S} I_{s,ij}\) term will monotonously increase.

Minimum redundancy maximum relevance

Minimum redundancy maximum relevance (mRMR) feature selection is similar to MIFS but differs in the way redundancy is used during optimisation (Peng, Long, and Ding 2005). Whereas for MIFS the optimisation criterion is \(I_{y,j} - \sum_{i\in S} I_{s,ij}\), in mRMR the optimisation criterion is \(I_{y,j} - \frac{1} {\left| S \right|} \sum_{i\in S} I_{s,ij}\), with \(\left| S \right|\) the number of features already selected.

Unlike in MIFS, the \(\frac{1}{\left|S\right|}\sum_{i\in S}I_{s,ij}\) term is not monotonically increasing. Consequently, features cannot be safely filtered. To limit computational complexity, we still remove features for which \(I_{y,j} - \frac{1} {\left| S \right| + 3} \sum_{i\in S} I_{s,ij} \leq 0\), as such features are unlikely to be selected.

Univariate and multivariate regression methods

Univariate and multivariate regression perform feature selection by performing regression using a feature or set of features as predictors. The performance of the regression model is then measured using a metric. Training and testing of regression models are repeated multiple times using bootstraps. For each bootstrap, the in-bag samples are used for training and the out-of-bag samples are using for testing.

This also defines the parameters of both methods, which are shown in the table below.

parameter	tag	values	optimised	comments
regression learner	`learner`	dependent on outcome	no	Any generalised linear regression model from the learning algorithms and hyperparameter optimisation vignette can be selected. Default values are `glm_logistic` for binomial, `glm_multinomial` for multinomial, `glm_gaussian` for continuous, `glm_poisson` for count, and `cox` for survival outcomes.
performance metric	`metric`	dependent on outcome	no	Any metric from the performance metrics vignette can be selected. Default values are `auc_roc` for binomial and multinomial, `mse` for continuous, `msle` for count and `concordance_index` for survival outcomes
number of bootstraps	`n_bootstrap`	\(\mathbb{Z} \in \left[1, \infty\right)\)	no	The default value is \(10\).
drop-out alpha level	`alpha`	\(\mathbb{R} \in \left[0, 1\right]\)	no	The default value is \(0.05\). Only used in multivariate regression.

Univariate regression

In the univariate regression method, a regression model is built with each feature separately using the in-bag data of the bootstrap. Then this model is evaluated using the metric, expressed using an objective representation (see computing the objective score in the learning algorithms and hyperparameter optimisation vignette). The objective representation \(s^*\) is computed on both in-bag (IB) and out-of-bag (OOB) data. Subsequently the balanced objective score \(f\) is computed: \(f=s^*_{OOB} - \left|s^*_{OOB}-s^*_{IB}\right|\).

The objective score \(f\) is subsequently averaged over all bootstraps to obtain the variable importance of a feature.

Multivariate regression

The procedure described for univariate regression forms the first step in multivariate regression. The rest follows forward selection. The most important feature is assigned to the subset of selected features and removed from the set of available features. Separate regression models are then built with each remaining feature and all the feature(s) in the selected feature subset as predictors. Thus, the subset of selected features iteratively increases in size until no features are remaining or the objective score no longer increases.

To limit mostly redundant computation, features that are unlikely to be selected are actively removed. To do so, the standard deviation of the objective score over the bootstraps is computed for each feature. The (one-sided, upper-tail) quantile \(q\) corresponding to the alpha-level indicated by parameter alpha is subsequently computed. If the obtained mean objective score is \(q\) standard deviations or more below the best objective score, the feature is removed.

Lasso, ridge and elastic net regression

Penalised regression is also a form of feature selection, as it selects an ‘optimal’ set of features to create a regression model. As features are usually normalised as part of pre-processing, the magnitude of each coefficient can be interpreted as its importance. All three shrinkage methods are implemented using the glmnet package (Hastie, Tibshirani, and Friedman 2009; Simon et al. 2011).

Only elastic net regression has a model hyperparameter that requires optimisation, but other parameters may be set as well, as shown in the table below:

parameter	tag	values	optimised	comments
family	`family`	`gaussian`, `binomial`, `poisson`, `multinomial`, `cox`	continuous outcomes	For continuous outcomes `gaussian` and `poisson` may be tested. The family is not optimised when it is specified, e.g. `lasso_gaussian`. For other outcomes only one applicable family exists.
elastic net penalty	`alpha`	\(\mathbb{R} \in \left[0,1\right]\)	elastic net	This penalty is fixed for ridge regression (`alpha = 0`) and lasso (`alpha = 1`).
optimal lambda	`lambda_min`	`lambda.1se`, `lambda.min`	no	Default is `lambda.min`.
number of CV folds	`n_folds`	\(\mathbb{Z} \in \left[3,n\right]\)	no	Default is \(3\) if \(n<30\), \(\lfloor n/10\rfloor\) if \(30\leq n \leq 200\) and \(20\) if \(n>200\).
normalisation	`normalise`	`FALSE`, `TRUE`	no	Default is `FALSE`, as normalisation is part of pre-processing in familiar.

Random forest-based methods

Several feature selection methods are based on random forests. All these methods require that a random forest model exists. Hence, familiar will train a random forest based on the training data. Random forest learners have a set of hyperparameters that are optimised prior to training, and these make up most of the method-specific parameters. These parameters, which are slightly different for ranger-based and randomForestSRC-based methods, are shown below.

parameter	tag	values	optimised	comments
number of trees	`n_tree`	\(\mathbb{Z} \in \left[0,\infty\right)\)	yes	This parameter is expressed on the \(\log_{2}\) scale, i.e. the actual input value will be \(2^\texttt{n_tree}\) (Oshiro, Perez, and Baranauskas 2012). The default range is \(\left[4, 10\right]\).
subsampling fraction	`sample_size`	\(\mathbb{R} \in \left(0, 1.0\right]\)	yes	Fraction of available data that is used for to create a single tree. The default range is \(\left[2 / m, 1.0\right]\), with \(m\) the number of samples.
number of features at each node	`m_try`	\(\mathbb{R} \in \left[0.0, 1.0\right]\)	yes	Familiar ensures that there is always at least one candidate feature.
node size	`node_size`	\(\mathbb{Z} \in \left[1, \infty\right)\)	yes	Minimum number of unique samples in terminal nodes. The default range is \(\left[5, \lfloor m / 3\rfloor\right]\), with \(m\) the number of samples.
maximum tree depth	`tree_depth`	\(\mathbb{Z} \in \left[1,\infty\right)\)	yes	Maximum depth to which trees are allowed to grow. The default range is \(\left[1, 10\right]\).
number of split points	`n_split`	\(\mathbb{Z} \in \left[0, \infty\right)\)	no	By default, splitting is deterministic and has one split point (\(0\)).
splitting rule (`randomForestSRC` only)	`split_rule`	`gini`, `auc`, `entropy`, `mse`, `quantile.regr`, `la.quantile.regr`, `logrank`, `logrankscore`, `bs.gradient`	no	Default splitting rules are `gini` for `binomial` and `multinonial` outcomes, `mse` for `continuous` and `count` outcomes and `logrank` for `survival` outcomes.
splitting rule (`ranger` only)	`split_rule`	`gini`, `hellinger`, `extratrees`, `beta`, `variance`, `logrank`, `C`, `maxstat`	no	Default splitting rules are `gini` for `binomial` and `multinomial` outcomes and `maxstat` for `continuous`, `count` and `survival` outcomes.
significance split threshold (`ranger` only)	`alpha`	\(\mathbb{R} \in \left(0.0, 1.0\right]\)	`maxstat`	Minimum significance level for further splitting. The default range is \(\left[10^{-6}, 1.0\right]\)
variable hunting cross-validation folds	`fs_vh_fold`	\(\mathbb{Z} \in \left[2, \infty\right)\)	no	Number of cross-validation folds for the `random_forest_variable_hunting` method. The default is \(5\).
variable hunting step size	`fs_vh_step_size`	\(\mathbb{Z} \in \left[1, \infty\right)\)	no	Step size for the `random_forest_variable_hunting` method. The default is \(1\).
variable hunting iterations	`fs_vh_n_rep`	\(\mathbb{Z} \in \left[1, \infty\right)\)	no	Number of Monte Carlo iterations for the `random_forest_variable_hunting` method. The default is \(50\).

The unoptimised methods do not require hyperparameter optimisation, and use default values from the ranger and randomForestSRC. For random_forest_variable_hunting_defaultthe fs_vh_fold, fs_vh_step_size and fs_vh_n_rep parameters can be set.

Permutation importance

The permutation importance method is implemented by random_forest_permutation and random_forest_permutation_default (randomForestSRC package) and random_forest_ranger_permutation and random_forest_ranger_permutation_default (ranger package). In short, this method functions as follows [Ishwaran2007-va]. As usual, each tree in the random forest is constructed using the in-bag samples of a bootstrap of the data. The predictive performance of each model is first measured using the out-of-bag data. Subsequently, the out-of-bag instances for each feature are randomly permuted, and predictive performance is assessed again. The difference between the normal performance and the permuted performance is used as a measure of the variable importance. For important features, this difference is large, whereas for irrelevant features the difference is negligible or even negative.

Holdout permutation importance

This variant on permutation importance (random_forest_ranger_holdout_permutation and random_forest_ranger_holdout_permutation_default) is implemented using ranger::holdoutRF. Instead of using out-of-bag to compute feature importance, two cross-validation folds are used. A random forest is trained on either fold, and variable importance determined on the other (Janitza, Celik, and Boulesteix 2018).

The hold-out variable importance method implemented in the randomForestSRC package (random_forest_holdout and random_forest_holdout_default) is implemented using randomForestSRC::holdout.vimp. It is similar to the previous variant, but does not cross-validation folds. Instead, out-of-bag prediction errors for models trained with and without each feature are compared.

Minimum depth variable selection

Important features tend to appear closer to the root of trees in random forests. Therefore, the position of each feature within a tree is assessed in minimum depth variable selection (Ishwaran et al. 2010).

Variable hunting

Variable hunting is implemented using the variable hunting algorithm implemented in randomForestSRC. Ishwaran suggest using it when minimum depth variable selection leads to high computational load, or a larger set of variables should be found (Ishwaran et al. 2010).

The variable hunting selection method has several parameters which can be set.

Impurity importance

At each node, the data is split into (two) subsets, which connects to two branches. After splitting, each single subset is purer than the parent dataset. As a concrete example, in regression problems the variance of each of the subsets is lower than that of the data prior to splitting. The decrease in variance specifically, or the decrease of impurity generally, is then used to assess feature importance.

familiar uses the impurity_corrected importance measure, which is unbiased to the number of split points of a feature and its distribution (Nembrini, König, and Wright 2018).

Special methods

Familiar offers several methods that are special in that they are not feature selection methods in the sense that they determine a variable importance that can be used for establishing feature rankings.

No feature selection

As the name suggests, the none method avoids feature selection altogether. All features are passed into a model. Feature order is randomly shuffled prior to building a model to avoid influence of the provided feature order.

Random feature selection

The random method randomly draws features prior to model building. It does not assign a random variable importance to a feature. New features are drawn each time a model is built. All features are available for the draw, but only \(m\) features are drawn. Here \(m\) is the signature size that is usually optimised by hyperparameter optimisation.

Signature only

When configuring familiar, any number of features can be set as a model signature using the signature configuration parameter. However, more features may be added to this signature through feature selection. To make sure that only the provided features enter a model, the signature_only method may be used.

Aggregating variable importance

In case of feature selection or modelling in the presence of resampling (e.g. bootstraps), the ranks of features may need to be aggregated across the different instances (Wald et al. 2012). The rank aggregation methods shown in the table below can be used for this purpose. Several methods require a threshold to indicate the size of the set of most highly ranked features, which can be set by specifying the vimp_aggregation_rank_threshold configuration parameter.

aggregation method	tag	comments
none	`none`
mean rank	`mean`
median rank	`median`
best rank	`best`
worst rank	`worst`
stability selection	`stability`	uses threshold
exponential selection	`exponential`	uses threshold
borda ranking	`borda`
enhanced borda ranking	`enhanced_borda`	uses threshold
truncated borda ranking	`truncated_borda`	uses threshold
enhanced truncated borda ranking	`enhanced_truncated_borda`	uses threshold

Notation

Let \(N\) be the number of ranking experiments that should be aggregated. Feature \(i\) for experiment \(j\) of \(N\) then has rank \(r_{ij}\). A lower rank indicates a more important feature. Some features may not receive a score during a ranking experiment, for example for multivariate variable importance methods such as lasso regression, or by use of a threshold \(\tau\). This is designated by \(\delta_{ij}\), which is \(0\) if the feature is absent, and \(1\) if it is present.

In case a threshold is used, \(\delta_{ij} = 1\) if \(r_{ij} \leq \tau\), and \(0\) otherwise.

Thus, for each experiment \(m_j = \sum^M_{i=1} \delta_{ij}\) features are ranked, out of \(M\) features. \(m_j\) is then also the maximum rank found in experiment \(j\).

Aggregating ranks for each feature results in an aggregate rank score \(s_i\). Features are subsequently ranked according to this method-specific score to arrive at an aggregate feature rank \(r_i\).

No rank aggregation

The none option does not aggregate ranks. Rather, scores are aggregated by computing the average score of a feature over all experiments that contain it. Ranks are then computed from the aggregated scores.

Mean rank aggregation

The mean rank aggregation method ranks features by computing the mean rank of a feature across all experiments that contain it.

\[s_i = \frac{\sum^{N}_{j=1} \delta_{ij} r_{ij}}{\sum^{N}_{j=1} \delta_{ij}}\]

The aggregate rank of features is then determined by sorting aggregate scores \(s_i\) in ascending order.

Median rank aggregation

The median rank aggregation method ranks features by computing the median rank of a feature across all experiments that contain it.

\[s_i = \underset{j \in N, \, \delta_{ij}=1}{\textrm{median}}(r_{ij})\]

The aggregate rank of features is then determined by sorting aggregate scores \(s_i\) in ascending order.

Best rank aggregation

The best rank aggregation method ranks features by the best rank that a feature has across all experiments that contain it.

\[s_i = \underset{j \in N, \, \delta_{ij}=1}{\textrm{min}} (r_{ij})\]

The aggregate rank of features is then determined by sorting aggregate scores \(s_i\) in ascending order.

Worst rank aggregation

The worst rank aggregation method ranks features by the worst rank that a feature has across all instances that contain it.

\[s_i = \underset{j \in N, \, \delta_{ij}=1}{\textrm{max}} (r_{ij})\]

The aggregate rank of features is then determined by sorting aggregate scores \(s_i\) in ascending order.

Stability rank aggregation

The stability aggregation method ranks features by their occurrence within the set of highly ranked features across all experiments. Our implementation generalises the method originally proposed by Meinshausen and Bühlmann (Meinshausen and Bühlmann 2010).

This method uses threshold \(\tau\) to designate the highly ranked features. Thus \(\delta_{ij} = 1\) if \(r_{ij} \leq \tau\), and \(0\) otherwise.

The aggregate rank score is computed as:

\[s_i = \frac{1}{N} \sum^N_{j=1} \delta_{ij}\]

The aggregate rank of features is then determined by sorting aggregate scores \(s_i\) in descending order, as more commonly occurring features are considered more important.

Exponential rank aggregation

The exponential aggregation method ranks features by the sum of the negative exponentials of their normalised ranks in instances where they occur within the set of highly ranked features. This method was originally suggested by Haury et al. (Haury, Gestraud, and Vert 2011).

This method uses threshold \(\tau\) to designate the highly ranked features. Thus \(\delta_{ij} = 1\) if \(r_{ij} \leq \tau\), and \(0\) otherwise.

\[s_i = \sum^N_{j=1} \delta_{ij} \exp({-r_{ij} / \tau)}\]