Package 'praznik'

Tools for information-based feature selection and scoring

Description

Praznik is a collection of tools for information theory-based feature selection and scoring.

Details

The first part of the package consists of efficient implementations of several popular information filters, feature selection approaches based on greedy optimisation of certain feature inclusion criterion. In a nutshell, an algorithm of this class requires an information system $(X,Y)$ and a predefined number of features to selected $k$, and works like this. To start, it estimates mutual information between each feature and the decision, find a feature with maximal such score and stores it as a first on a list of selected features, $S$. Then, it estimates a value of a certain criterion function $J(X,Y,S)$ for each feature $X$; this function also depends on $Y$ and the set of already selected features $S$. As in the first step, the previously unselected feature with a greatest value of the criterion function is selected next. This is repeated until the method would gather $k$ features, or, in case of some methods, when no more informative features can be found. The methods implemented in praznik consider the following criteria.

The mutual information maximisation filter, MIM, simply selects top-$k$ features of best mutual information, that is

$J_{MIM}=I(X;Y).$

The minimal conditional mutual information maximisation proposed by F. Fleauret, CMIM, uses

$J_{CMIM}(X)=\min(I(X;Y),\min_{W\in S} I(X;Y|W));$

this method is also effectively identical to the information fragments method.

The minimum redundancy maximal relevancy proposed by H. Peng et al., MRMR, uses

$J_{MRMR}=I(X;Y)-\frac{1}{|S|}\sum_{W\in S} I(X;W).$

The joint mutual information filter by H. Yang and J. Moody, JMI, uses

$J_{JMI}=\sum_{W\in S} I(X,W;Y).$

The double input symmetrical relevance filter by P. Meyer and G. Bontempi, DISR, uses

$J_{DISR}(X)=\sum_{W\in S} \frac{I(X,W;Y)}{H(X,W,Y)}.$

The minimal joint mutual information maximisation filter by M. Bennasar, Y. Hicks and R. Setchi, JMIM, uses

$J_{JMIM}=\min_{W\in S} I(X,W;Y).$

The minimal normalised joint mutual information maximisation filter by the same authors, NJMIM, uses

$J_{NJMIM}=\min_{W\in S} \frac{I(X,W;Y)}{H(X,W,Y)}.$

The third-order joint mutual information filter by Sechidis et al., JMI3, uses

$J(X)=\frac{1}{2}\sum_{(U,W)\in S^2; U\neq W} I(X,U,W;Y).$

While CMIM, JMIM and NJMIM consider minimal value over already selected features, they may use a somewhat more sophisticated and faster algorithm.

The second part of the package provides methods for scoring features, useful on its own as well as building blocks for more sophisticated algorithms. In particular, the package exposes the following functions:

hScores returns

$H(X).$

jhScores returns

$H(X,Y).$

miScores returns

$I(X;Y).$

cmiScores returns, for a given condition vector $Z$,

$I(X;Y|Z).$

jmiScores returns

$I(X,Z;Y).$

njmiScores returns

$\frac{I(X,Z;Y)}{H(X,Y,Z)}.$

minCmiScores, maxCmiScores and minMaxCmiScores return

$\min_j I(X_i;Y|X_j)$

and/or

$\max_j I(X_i;Y|X_j).$

maxJmiScores returns

$\max_{j\neq i} I(X_i,X_j;Y).$

triScores returns, for every triple of features,

$I(X_i;X_j;X_k).$

These function generally also have their *Matrix counterparts, which efficiently build a matrix of scores between all pairs of features. This is especially useful for network inference tasks.

Estimation of mutual information and its generalisations is a hard task; still, praznik aims at speed and simplicity and hence only offers basic, maximum likelihood estimator applicable on discrete data. For convenience, praznik automatically and silently coerces non-factor inputs into about ten equally-spaced bins, following the heuristic often used in literature.

Furthermore, praznik provides kTransform function for converting continuous features into discrete ones with Kendall transformation, a novel approach based on Kendall correlation coefficient which allows for multivariate reasoning based on monotonicity agreement.

Additionally, praznik has a limited, experimental support for replacing entropic statistics with Gini impurity-based; in such framework, entropy is replaced by Gini impurity

$g(X):=1-\sum_x p_x^2,$

which leads to an impurity gain

$G(X;Y):=g(Y)-E(g(Y)|X)=\sum_{xy}\frac{p_{xy}^2}{p_x}-\sum_{y} p_y^2,$

a counterpart of mutual information or information gain. It does not possess most of elegant properties of mutual information, yet values of both are usually highly correlated; moreover, Gini gain is computationally easier to calculate, hence it often replaces MI in performance-sensitive applications, like optimising splits in decision trees.

In a present version, praznik includes impScores for generating values of $G$ for all features (an analog of miScores, as well as JIM, a Gini gain-based feature selection method otherwise identical to JMI.

Author(s)

Maintainer: Miron B. Kursa [email protected] (ORCID)

References

"Praznik: High performance information-based feature selection" M.B. Kursa. SoftwareX 16, 100819 (2021).

"Conditional Likelihood Maximisation: A Unifying Framework for Information Theoretic Feature Selection" G. Brown et al. JMLR (2012).

See Also

Useful links:

• https://gitlab.com/mbq/praznik

• Report bugs at https://gitlab.com/mbq/praznik/-/issues

---

Conditional mutual information maximisation filter

Description

The method starts with a feature of a maximal mutual information with the decision $Y$. Then, it greedily adds feature $X$ with a maximal value of the following criterion:

$J(X)=I(X;Y|S),$

where $S$ is the set of already selected features.

Usage

CMI(X, Y, k = 3, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
Y	Decision attribute; should be given as a factor, but other options are accepted, exactly like for attributes. NAs are not allowed.
k	Number of attributes to select. Must not exceed ncol(X).
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A list with two elements: selection, a vector of indices of the selected features in the selection order, and score, a vector of corresponding feature scores. Names of both vectors will correspond to the names of features in X. Both vectors will be at most of a length k, as the selection may stop sooner, even during initial selection, in which case both vectors will be empty.

Note

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

Examples

data(MadelonD)
CMI(MadelonD$X,MadelonD$Y,20)

---

Minimal conditional mutual information maximisation filter

Description

The method starts with a feature of a maximal mutual information with the decision $Y$. Then, it greedily adds feature $X$ with a maximal value of the following criterion:

$J(X)=\min(I(X;Y),\min_{W\in S} I(X;Y|W)),$

where $S$ is the set of already selected features.

Usage

CMIM(X, Y, k = 3, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
Y	Decision attribute; should be given as a factor, but other options are accepted, exactly like for attributes. NAs are not allowed.
k	Number of attributes to select. Must not exceed ncol(X).
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A list with two elements: selection, a vector of indices of the selected features in the selection order, and score, a vector of corresponding feature scores. Names of both vectors will correspond to the names of features in X. Both vectors will be at most of a length k, as the selection may stop sooner, even during initial selection, in which case both vectors will be empty.

Note

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

References

"Fast Binary Feature Selection using Conditional Mutual Information Maximisation" F. Fleuret, JMLR (2004)

"Object recognition with informative features and linear classification" M. Vidal-Naquet and S. Ullman, IEEE Conference on Computer Vision and Pattern Recognition (2003).

Examples

data(MadelonD)
CMIM(MadelonD$X,MadelonD$Y,20)

---

Conditional mutual information matrix with a common condition

Description

Calculates conditional mutual information between each two features given another one, that is

$I(X_i;X_j|Z).$

Usage

cmiMatrix(X, Z, zeroDiag = TRUE, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
Z	Condition; should be given as a factor, but other options are accepted, as for features.
zeroDiag	Boolean flag, whether the diagonal should be filled with zeroes, or with degenerated scores for two identical copies of a feature.
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A numerical matrix with scores, with row and column names copied from X.

Note

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

Examples

cmiMatrix(iris[,-5],iris[,5])

---

Conditional mutual information scores

Description

Calculates conditional mutual information between each features and the decision, that is

$I(X;Y|Z).$

Usage

cmiScores(X, Y, Z, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
Y	Decision attribute; should be given as a factor, but other options are accepted, exactly like for attributes. NAs are not allowed.
Z	Condition; should be given as a factor, but other options are accepted, as for features.
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A numerical vector with conditional mutual information scores, with names copied from X.

Note

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

Examples

cmiScores(iris[,-5],iris$Species,iris$Sepal.Length)

---

Double input symmetrical relevance filter

Description

The method starts with a feature of a maximal mutual information with the decision $Y$. Then, it greedily adds feature $X$ with a maximal value of the following criterion:

$J(X)=\sum_{W\in S} \frac{I(X,W;Y)}{H(X,W,Y)},$

where $S$ is the set of already selected features.

Usage

DISR(X, Y, k = 3, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
Y	Decision attribute; should be given as a factor, but other options are accepted, exactly like for attributes. NAs are not allowed.
k	Number of attributes to select. Must not exceed ncol(X).
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A list with two elements: selection, a vector of indices of the selected features in the selection order, and score, a vector of corresponding feature scores. Names of both vectors will correspond to the names of features in X. Both vectors will be at most of a length k, as the selection may stop sooner, even during initial selection, in which case both vectors will be empty.

Note

DISR is a normalised version of JMI; JMIM and NJMIM are modifications of JMI and DISR in which minimal joint information over already selected features is used instead of a sum.

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

References

"On the Use of Variable Complementarity for Feature Selection in Cancer Classification" P. Meyer and G. Bontempi, (2006)

Examples

data(MadelonD)
DISR(MadelonD$X,MadelonD$Y,20)

---

Directional normalised mutual information matrix

Description

Calculates directed normalised mutual information between each two features, that is

$\frac{I(X_i,X_j)}{H(X_j)}.$

Usage

dnmiMatrix(X, zeroDiag = TRUE, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
zeroDiag	Boolean flag, whether the diagonal should be filled with zeroes, or with degenerated scores for two identical copies of a feature.
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A numerical matrix with scores, with row and column names copied from X.

Note

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

Examples

dnmiMatrix(iris)

---

Entropy scores

Description

Calculates entropy of each feature, that is

$H(X).$

Usage

hScores(X, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A numerical vector with entropy scores, with names copied from X.

Note

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

Examples

hScores(iris[,-5])

---

Conditional mutual information matrix with a common variable

Description

Calculates conditional mutual information between each feature and the decision given each other feature, that is

$I(X_i;Y|X_j).$

Usage

icmiMatrix(X, Y, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
Y	Decision attribute; should be given as a factor, but other options are accepted, exactly like for attributes. NAs are not allowed.
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A numerical matrix with scores, with row and column names copied from X.

Note

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

Diagonal is always zero with this score. The function name comes from the reasoning that this is an "interaction-CMI" showing how feature pairs interact in explaining the decision.

Examples

icmiMatrix(iris[,-5],iris[,5])

---

Gini impurity scores

Description

Calculates Gini impurity between each feature and the decision, that is

$G(X;Y)=\sum_{xy} \frac{p_{xy}^2}{p_x}-\sum_y p_y^2.$

Usage

impScores(X, Y, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
Y	Decision attribute; should be given as a factor, but other options are accepted, exactly like for attributes. NAs are not allowed.
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A numerical vector with Gini impurity scores, with names copied from X.

Note

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

Examples

impScores(iris[,-5],iris$Species)

---

Joint entropy scores

Description

Calculates joint entropy of each feature and a condition Y, that is

$H(X,Y).$

Usage

jhScores(X, Y, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
Y	Decision attribute; should be given as a factor, but other options are accepted, exactly like for attributes. NAs are not allowed.
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A numerical vector with entropy scores, with names copied from X.

Note

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

Examples

jhScores(iris[,-5],iris[,5])

---

Joint impurity filter

Description

The method starts with a feature of a maximal impurity gain with the decision $Y$. Then, it greedily adds feature $X$ with a maximal value of the following criterion:

$J(X)=\sum_{W\in S} G(X,W;Y),$

where $S$ is the set of already selected features, and

$G(X;Y)=\sum_{xy}\frac{p_{xy}^2}{p_x}-\sum_{y} p_y^2$

is the Gini impurity gain from partitioning $Y$ according to $X$.

Usage

JIM(X, Y, k = 3, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
Y	Decision attribute; should be given as a factor, but other options are accepted, exactly like for attributes. NAs are not allowed.
k	Number of attributes to select. Must not exceed ncol(X).
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A list with two elements: selection, a vector of indices of the selected features in the selection order, and score, a vector of corresponding feature scores. Names of both vectors will correspond to the names of features in X. Both vectors will be at most of a length k, as the selection may stop sooner, even during initial selection, in which case both vectors will be empty.

Note

This is an impurity-based version of JMI; expect similar results in slightly shorter time.

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

Examples

data(MadelonD)
JIM(MadelonD$X,MadelonD$Y,20)

---

Joint mutual information filter

Description

The method starts with a feature of a maximal mutual information with the decision $Y$. Then, it greedily adds feature $X$ with a maximal value of the following criterion:

$J(X)=\sum_{W\in S} I(X,W;Y),$

where $S$ is the set of already selected features.

Usage

JMI(X, Y, k = 3, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
Y	Decision attribute; should be given as a factor, but other options are accepted, exactly like for attributes. NAs are not allowed.
k	Number of attributes to select. Must not exceed ncol(X).
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A list with two elements: selection, a vector of indices of the selected features in the selection order, and score, a vector of corresponding feature scores. Names of both vectors will correspond to the names of features in X. Both vectors will be at most of a length k, as the selection may stop sooner, even during initial selection, in which case both vectors will be empty.

Note

DISR is a normalised version of JMI; JMIM and NJMIM are modifications of JMI and DISR in which minimal joint information over already selected features is used instead of a sum.

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

References

"Data Visualization and Feature Selection: New Algorithms for Nongaussian Data" H. Yang and J. Moody, NIPS (1999)

Examples

data(MadelonD)
JMI(MadelonD$X,MadelonD$Y,20)

---

Third-order joint mutual information filter

Description

The method starts with two features: $X_1$ of a maximal mutual information with the decision $Y$, and $X_2$ of a maximal value of $I(X_1,X_2;Y)$, as would be selected second by a regular JMI. Then, it greedily adds feature $X$ with a maximal value of the following criterion:

$J(X)=\frac{1}{2}\sum_{(U,W)\in S^2; U\neq W} I(X,U,W;Y),$

where $S$ is the set of already selected features.

Usage

JMI3(X, Y, k = 3, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
Y	Decision attribute; should be given as a factor, but other options are accepted, exactly like for attributes. NAs are not allowed.
k	Number of attributes to select. Must not exceed ncol(X).
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A list with two elements: selection, a vector of indices of the selected features in the selection order, and score, a vector of corresponding feature scores. Names of both vectors will correspond to the names of features in X. Both vectors will be at most of a length k, as the selection may stop sooner, even during initial selection, in which case both vectors will be empty.

Note

This method has a complexity of $O(k^2\cdot m \cdot n)$, while other filters have $O(k\cdot m \cdot n)$ — for larger $k$, it will be substantially slower. In the original paper, special shrinkage estimator of MI is used; in praznik, all algorithms use ML estimators, so is JMI3.

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

References

"Efficient feature selection using shrinkage estimators" K. Sechidis, L. Azzimonti, A. Pocock, G. Corani, J. Weatherall and G. Brown. Machine Learning, 108 (8-9), pp. 1261-1286 (2019)

Examples

## Not run: data(MadelonD)
JMI3(MadelonD$X,MadelonD$Y,20)
## End(Not run)

---

Minimal joint mutual information maximisation filter

Description

The method starts with a feature of a maximal mutual information with the decision $Y$. Then, it greedily adds feature $X$ with a maximal value of the following criterion:

$J(X)=\min_{W\in S} I(X,W;Y),$

where $S$ is the set of already selected features.

Usage

JMIM(X, Y, k = 3, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
Y	Decision attribute; should be given as a factor, but other options are accepted, exactly like for attributes. NAs are not allowed.
k	Number of attributes to select. Must not exceed ncol(X).
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A list with two elements: selection, a vector of indices of the selected features in the selection order, and score, a vector of corresponding feature scores. Names of both vectors will correspond to the names of features in X. Both vectors will be at most of a length k, as the selection may stop sooner, even during initial selection, in which case both vectors will be empty.

Note

NJMIM is a normalised version of JMIM; JMI and DISR are modifications of JMIM and NJMIM in which a sum of joint information over already selected features is used instead of a minimum.

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

References

"Feature selection using Joint Mutual Information Maximisation" M. Bennasar, Y. Hicks and R. Setchi, (2015)

Examples

data(MadelonD)
JMIM(MadelonD$X,MadelonD$Y,20)

---

Joint mutual information matrix

Description

Calculates mutual information between each feature and a joint mix of each other feature with a given feature, that is

$I(X_i;X_j,Z).$

Usage

jmiMatrix(X, Z, zeroDiag = TRUE, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
Z	Condition; should be given as a factor, but other options are accepted, as for features.
zeroDiag	Boolean flag, whether the diagonal should be filled with zeroes, or with degenerated scores for two identical copies of a feature.
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A numerical matrix with scores, with row and column names copied from X.

Note

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

Examples

jmiMatrix(iris[,-5],iris[,5])

---

Joint mutual information scores

Description

Calculates joint mutual information between each feature joint with some other vector Z with the decision, that is

$I(X,Z;Y).$

This is the same as conditional mutual information between X and Y plus a constant that depends on Y and Z, that is

$I(X,Z;Y)=I(X;Y|Z)+I(Y;Z).$

Usage

jmiScores(X, Y, Z, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
Y	Decision attribute; should be given as a factor, but other options are accepted, exactly like for attributes. NAs are not allowed.
Z	Other vector; should be given as a factor, but other options are accepted, as for features.
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A numerical vector with joint mutual information scores, with names copied from X.

Note

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

Examples

jmiScores(iris[,-5],iris$Species,iris$Sepal.Length)

---

Join factors

Description

Convenience function for joining factors.

Usage

joinf(...)

Arguments

...	One or more features to merge. Given as single vectors or data.frames. Accepted feature types are factor (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation). NAs are not allowed.

Value

Joint factor, with levels l1 to l<n>. Vacant combinations are dropped.

Note

You can pass a single vector to this function to see how praznik interprets it.

Examples

joinf(c(1,2,1,2),c(1,1,2,2))

---

Inverse Kendall transform

Description

This function attempts to reverse Kendall transformation using a simple ranking agreement method, which always restores original ranking if the input corresponds to one, or a reasonable best-effort guess if not. Namely, each objects gets a score based on its relation with each other object, 2 points for a win ('>') and 1 point for a tie ('='); these scores are used to calculate ranks. This function can also be directly given greater-than scores, for instance confidence scores from some classifier trained on Kendall-transformed data.

Usage

kInverse(x)

Arguments

x	A Kendall-transformed feature to be converted back into a ranking. To be interpreted as a such, it must be a factor with levels being a subset of '<', '>' or '='. Alternatively, it may be a numeric vector of greater-than scores.

Value

Vector of ranks corresponding to x.

Note

An order of elements in x is crucial; if it is not the same as generated by the kTransform, results will be wrong. This function cannot assert that the order is correct.

References

"Kendall transformation brings a robust categorical representation of ordinal data" M.B. Kursa. SciRep 12, 8341 (2022).

Examples

kInverse(kTransform(1:7))

---

Kendall transformation

Description

Kendall transformation

Usage

kTransform(x)

Arguments

x	Vector or data frame to be Kendall-transformed; allowed feature types are numeric, integer (treated as numeric), ordered factor, logical and unordered factor with two or less levels. NA and non-finite values are allowed; NaN is treated as NA.

Value

A transformed vector or data frame with transformed columns.

References

"Kendall transformation brings a robust categorical representation of ordinal data" M.B. Kursa. SciRep 12, 8341 (2022).

Examples

kTransform(data.frame(Asc=1:3,Desc=3:1,Vsh=c(2,1,2)))

---

Pre-discretised Madelon dataset

Description

Madelon is a synthetic data set from the NIPS 2003 feature selection challenge, generated by Isabelle Guyon. It contains 480 irrelevant and 20 relevant features, including 5 informative and 15 redundant. In this version, the originally numerical features have been pre-cut into 10 bins, as well as their names have been altered to reveal 20 relevant features (as identified by the Boruta method).

Usage

data(MadelonD)

Format

A list with two elements, X containing a data frame with predictors, and Y, the decision. Features are in the same order as in the original data; the names of relevant ones start with Rel, while of irrelevant ones with Irr.

Source

https://archive.ics.uci.edu/ml/datasets/Madelon

---

Maximal pairwise conditional mutual information scores

Description

For each feature, calculates the conditional mutual information between this feature and the decision, conditioned on all other features, and returns the maximal value, that is

$max_j I(X_i;Y|X_j).$

Usage

maxCmiScores(X, Y, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
Y	Decision attribute; should be given as a factor, but other options are accepted, exactly like for attributes. NAs are not allowed.
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A numerical vector with maximal pairwise conditional mutual information scores, with names copied from X.

Note

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

Examples

maxCmiScores(iris[,-5],iris$Species)

---

Maximal pairwise joint mutual information scores

Description

Calculates joint mutual information between each joint feature pair with the decision, and yields maximal value for each feature, that is

$max_{j\neq i} I(X_i,X_j;Y).$

Usage

maxJmiScores(X, Y, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
Y	Decision attribute; should be given as a factor, but other options are accepted, exactly like for attributes. NAs are not allowed.
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A numerical vector with maximal pairwise joint mutual information scores, with names copied from X.

Note

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

Examples

maxJmiScores(iris[,-5],iris$Species)

---

Mutual information maximisation filter

Description

Calculates mutual information between all features and the decision, then returns top k.

Usage

MIM(X, Y, k = 3, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
Y	Decision attribute; should be given as a factor, but other options are accepted, exactly like for attributes. NAs are not allowed.
k	Number of attributes to select. Must not exceed ncol(X).
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A list with two elements: selection, a vector of indices of the selected features in the selection order, and score, a vector of corresponding feature scores. Names of both vectors will correspond to the names of features in X. Both vectors will be at most of a length k, as the selection may stop sooner, even during initial selection, in which case both vectors will be empty.

Note

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

Examples

data(MadelonD)
MIM(MadelonD$X,MadelonD$Y,20)

---

Mutual information matrix

Description

Calculates mutual information between each two features, that is

$I(X_i,X_j).$

Usage

miMatrix(X, zeroDiag = TRUE, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
zeroDiag	Boolean flag, whether the diagonal should be filled with zeroes, or with degenerated scores for two identical copies of a feature.
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A numerical matrix with scores, with row and column names copied from X.

Note

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

Examples

miMatrix(iris)

---

Minimal pairwise conditional mutual information scores

Description

For each feature, calculates the conditional mutual information between this feature and the decision, conditioned on all other features, and returns the minimal value, that is

$min_j I(X_i;Y|X_j).$

Usage

minCmiScores(X, Y, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
Y	Decision attribute; should be given as a factor, but other options are accepted, exactly like for attributes. NAs are not allowed.
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A numerical vector with minimal pairwise conditional mutual information scores, with names copied from X.

Note

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

Examples

minCmiScores(iris[,-5],iris$Species)

---

Extreme values of pairwise conditional mutual information scores

Description

For each feature, calculates the conditional mutual information between this feature and the decision, conditioned on all other features, and returns extreme values, that is

$min_j I(X_i;Y|X_j)$

and

$max_j I(X_i;Y|X_j).$

Usage

minMaxCmiScores(X, Y, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
Y	Decision attribute; should be given as a factor, but other options are accepted, exactly like for attributes. NAs are not allowed.
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A numerical matrix with minimal (first row) and maximal (second row) pairwise conditional mutual information scores, with names copied from X.

Note

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

Examples

minMaxCmiScores(iris[,-5],iris$Species)

---

Mutual information scores

Description

Calculates mutual information between each feature and the decision, that is

$I(X,Y).$

Usage

miScores(X, Y, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
Y	Decision attribute; should be given as a factor, but other options are accepted, exactly like for attributes. NAs are not allowed.
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A numerical vector with mutual information scores, with names copied from X.

Note

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

Examples

miScores(iris[,-5],iris$Species)

---

Minimum redundancy maximal relevancy filter

Description

The method starts with a feature of a maximal mutual information with the decision $Y$. Then, it greedily adds feature $X$ with a maximal value of the following criterion:

$J(X)=I(X;Y)-\frac{1}{|S|}\sum_{W\in S} I(X;W),$

where $S$ is the set of already selected features.

Usage

MRMR(X, Y, k = if (positive) ncol(X) else 3, positive = FALSE, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
Y	Decision attribute; should be given as a factor, but other options are accepted, exactly like for attributes. NAs are not allowed.
k	Number of attributes to select. Must not exceed ncol(X).
positive	If true, algorithm won't return features with negative scores (i.e., with redundancy term higher than the relevance therm). In that case, k controls the maximal number of returned features, and is set to 'ncol(X)' by default.
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A list with two elements: selection, a vector of indices of the selected features in the selection order, and score, a vector of corresponding feature scores. Names of both vectors will correspond to the names of features in X. Both vectors will be at most of a length k, as the selection may stop sooner, even during initial selection, in which case both vectors will be empty.

Note

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

References

"Feature Selection Based on Mutual Information: Criteria of Max-Dependency, Max-Relevance, and Min-Redundancy" H. Peng et al. IEEE Pattern Analysis and Machine Intelligence (PAMI) (2005)

Examples

data(MadelonD)
MRMR(MadelonD$X,MadelonD$Y,20)

---

Minimal normalised joint mutual information maximisation filter

Description

The method starts with a feature of a maximal mutual information with the decision $Y$. Then, it greedily adds feature $X$ with a maximal value of the following criterion:

$J(X)=\min_{W\in S} \frac{I(X,W;Y)}{H(X,W,Y)},$

where $S$ is the set of already selected features.

Usage

NJMIM(X, Y, k = 3, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
Y	Decision attribute; should be given as a factor, but other options are accepted, exactly like for attributes. NAs are not allowed.
k	Number of attributes to select. Must not exceed ncol(X).
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A list with two elements: selection, a vector of indices of the selected features in the selection order, and score, a vector of corresponding feature scores. Names of both vectors will correspond to the names of features in X. Both vectors will be at most of a length k, as the selection may stop sooner, even during initial selection, in which case both vectors will be empty.

Note

NJMIM is a normalised version of JMIM; JMI and DISR are modifications of JMIM and NJMIM in which a sum of joint information over already selected features is used instead of a minimum. It stops returning features when the best score reaches 0.

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

References

"Feature selection using Joint Mutual Information Maximisation" M. Bennasar, Y. Hicks and R. Setchi, (2015)

Examples

data(MadelonD)
NJMIM(MadelonD$X,MadelonD$Y,20)

---

Normalised joint mutual information matrix

Description

Calculates normalised mutual information between each feature and a joint mix of each other feature with a given feature, that is

$\frac{I(X_i;X_j,Z)}{H(X_i,X_j,Z)}.$

Usage

njmiMatrix(X, Z, zeroDiag = TRUE, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
Z	Condition; should be given as a factor, but other options are accepted, as for features.
zeroDiag	Boolean flag, whether the diagonal should be filled with zeroes, or with degenerated scores for two identical copies of a feature.
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A numerical matrix with scores, with row and column names copied from X.

Note

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

Examples

njmiMatrix(iris[,-5],iris[,5])

---

Normalised joint mutual information scores

Description

Calculated normalised mutual information between each feature joint with some other vector Z and the decision, that is

$\frac{I(X,Z;Y)}{H(X,Y,Z)}.$

This is the same as in the criterion used by DISR and NJMIM.

Usage

njmiScores(X, Y, Z, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
Y	Decision attribute; should be given as a factor, but other options are accepted, exactly like for attributes. NAs are not allowed.
Z	Other vector; should be given as a factor, but other options are accepted, as for features.
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A numerical vector with the normalised joint mutual information scores, with names copied from X.

Note

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

Examples

njmiScores(iris[,-5],iris$Species,iris$Sepal.Length)

---

Normalised mutual information matrix

Description

Calculates normalised mutual information between each two features, that is

$\frac{I(X_i,X_j)}{H(X_i,X_j)}.$

Usage

nmiMatrix(X, zeroDiag = TRUE, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
zeroDiag	Boolean flag, whether the diagonal should be filled with zeroes, or with degenerated scores for two identical copies of a feature.
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A numerical matrix with scores, with row and column names copied from X.

Note

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

Examples

nmiMatrix(iris)

---

Mutual information of feature triples

Description

Calculates mutual information of each triple of features, that is

$I(X_i;X_j;X_k).$

Usage

triScores(X, threads = 0)

Arguments

X	Attribute table, given as a data frame with either factors (preferred), booleans, integers (treated as categorical) or reals (which undergo automatic categorisation; see below for details). Single vector will be interpreted as a data.frame with one column. NAs are not allowed.
threads	Number of threads to use; default value, 0, means all available to OpenMP.

Value

A data frame with four columns; first three (Var1, Var2 and Var3) are names of features, fourth, MI is the value of the mutual information. The order of features does not matter, hence only

$n(n-1)(n-2)/6$

unique, sorted triples are evaluated.

Note

The method requires input to be discrete to use empirical estimators of distribution, and, consequently, information gain or entropy. To allow smoother user experience, praznik automatically coerces non-factor vectors in inputs, which requires additional time, memory and may yield confusing results – the best practice is to convert data to factors prior to feeding them in this function. Real attributes are cut into about 10 equally-spaced bins, following the heuristic often used in literature. Precise number of cuts depends on the number of objects; namely, it is $n/3$, but never less than 2 and never more than 10. Integers (which technically are also numeric) are treated as categorical variables (for compatibility with similar software), so in a very different way – one should be aware that an actually numeric attribute which happens to be an integer could be coerced into a $n$-level categorical, which would have a perfect mutual information score and would likely become a very disruptive false positive.

In a current version, the maximal number of features accepted is 2345, which gives a bit less than 2^32 triples. The equation used for calculation is

$I(X_i;X_j;X_k)=I(X_i;X_k)+I(X_j;X_k)-I(X_i,X_j;X_k).$

Henceforth, please mind that rounding errors may occur and influence reproducibility.

Examples

triScores(iris)

---