Skip to content

Latest commit

 

History

History
396 lines (246 loc) · 27.1 KB

README.md

File metadata and controls

396 lines (246 loc) · 27.1 KB

Rstats crates.io crates.io GitHub last commit Actions Status

Author: Libor Spacek

Usage

This crate is written in 100% safe Rust.

Use in your source files any of the following structs, as and when needed:

use Rstats::{RE,RError,Params,TriangMat,MinMax};

and any of the following traits:

use Rstats::{Stats,Vecg,Vecu8,MutVecg,VecVec,VecVecg};

and any of the following auxiliary functions:

use Rstats::{
    fromop,sumn,tm_stat,unit_matrix,nodata_error,data_error,
    arith_error,other_error };

or just simply use everything:

use Rstats::*;

The latest (nightly) version is always available in the github repository Rstats. Sometimes it may be (only in some details) a little ahead of the crates.io release versions.

It is highly recommended to read and run tests.rs for examples of usage. To run all the tests, use a single thread in order not to print the results in confusing mixed-up order:

cargo test --release -- --test-threads=1 --nocapture

However, geometric_medians, which compares multithreading performance, should be run separately in multiple threads, as follows:

cargo test -r geometric_medians -- --nocapture

Alternatively, just to get a quick idea of the methods provided and their usage, read the output produced by an automated test run. There are test logs generated for each new push to the github repository. Click the latest (top) one, then Rstats and then Run cargo test ... The badge at the top of this document lights up green when all the tests have passed and clicking it gets you to these logs as well.

Any compilation errors arising out of rstats crate indicate most likely that some of the dependencies are out of date. Issuing cargo update command will usually fix this.

Introduction

Rstats has a small footprint. Only the best methods are implemented, primarily with Data Analysis and Machine Learning in mind. They include multidimensional (nd or 'hyperspace') analysis, i.e. characterising clouds of n points in space of d dimensions.

Several branches of mathematics: statistics, information theory, set theory and linear algebra are combined in this one consistent crate, based on the abstraction that they all operate on the same data objects (here Rust Vecs). The only difference being that an ordering of their components is sometimes assumed (in linear algebra, set theory) and sometimes it is not (in statistics, information theory, set theory).

Rstats begins with basic statistical measures, information measures, vector algebra and linear algebra. These provide self-contained tools for the multidimensional algorithms but they are also useful in their own right.

Non analytical (non parametric) statistics is preferred, whereby the 'random variables' are replaced by vectors of real data. Probabilities densities and other parameters are in preference obtained from the real data (pivotal quantity), not from some assumed distributions.

Linear algebra uses generic data structure Vec<Vec<T>> capable of representing irregular matrices.

Struct TriangMat is defined and used for symmetric, anti-symmetric, and triangular matrices, and their transposed versions, saving memory.

Our treatment of multidimensional sets of points is constructed from the first principles. Some original concepts, not found elsewhere, are defined and implemented here (see the next section).

Zero median vectors are generally preferred to commonly used zero mean vectors.

In n dimensions, many authors 'cheat' by using quasi medians (one dimensional (1d) medians along each axis). Quasi medians are a poor start to stable characterisation of multidimensional data. Also, they are actually slower to compute than our gm ( geometric median), as soon as the number of dimensions exceeds trivial numbers.

Specifically, all such 1d measures are sensitive to the choice of axis and thus are affected by their rotation.

In contrast, our methods based on gm are axis (rotation) independent. Also, they are more stable, as medians have a maximum possible breakdown point.

We compute geometric medians by our method gmedian and its parallel version par_gmedian in trait VecVec and their weighted versions wgmedian and par_wgmedian in trait VecVecg. It is mostly these efficient algorithms that make our new concepts described below practical.

Additional Documentation

For more detailed comments, plus some examples, see rstats in docs.rs. You may have to go directly to the modules source. These traits are implemented for existing 'out of this crate' rust Vec type and unfortunately rust docs do not display 'implementations on foreign types' very well.

New Concepts and their Definitions

  • zero median points (or vectors) are obtained by moving the origin of the coordinate system to the median (in 1d), or to the gm (in nd). They are our alternative to the commonly used zero mean points, obtained by moving the origin to the arithmetic mean (in 1d) or to the arithmetic centroid (in nd).

  • median correlation between two 1d sets of the same length.
    We define this correlation similarly to Pearson, as cosine of an angle between two normalised samples, interpreted as coordinate vectors. Pearson first normalises each set by subtracting its mean from all components. Whereas we subtract the median (cf. zero median points above). This conceptual clarity is one of the benefits of interpreting a data sample of length d as a single vector in d dimensional space.

  • gmedian, par_gmedian, wgmedian and par_wgmedian
    our fast multidimensional geometric median (gm) algorithms.

  • madgm (median of distances from gm)
    is our generalisation of mad (median of absolute differences from median), to n dimensions. 1d median is replaced in nd by gm. Where mad was a robust measure of 1d data spread, madgm becomes a robust measure of nd data spread. We define it as: median(|pi-gm|,i=1..n), where p1..pn are a sample of n data points, each of which is now a vector.

  • tm_stat
    We define our generalized tm_stat of a single scalar observation x as: (x-centre)/spread, with the recommendation to replace mean by median and std by mad, whenever possible. Compare with common t-stat, defined as (x-mean)/std, where std is the standard deviation.
    These are similar to the well known standard z-score, except that the central tendency and spread are obtained from the sample (pivotal quantity), rather than from any old assumed population distribution.

  • tm_statistic
    we now generalize tm_stat from scalar domain to vector domain of any number of dimensions, defining tm_statistic as |p-gm|/madgm, where p is a single observation point in nd space. For sample central tendency now serves the geometric median gm vector and the spread is the madgm scalar (see above). The error distance of observation p from the median: |p-gm|, is also a scalar. Thus the co-domain of tm_statistic is a simple positive scalar, regardless of the dimensionality of the vector space in question.

  • contribution
    one of the key questions of Machine Learning is how to quantify the contribution that each example (typically represented as a member of some large nd set) makes to the recognition concept, or outcome class, represented by that set. In answer to this, we define the contribution of a point p as the magnitude of displacement of gm, caused by adding p to the set. Generally, outlying points make greater contributions to the gm but not as much as to the centroid. The contribution depends not only on the radius of p but also on the radii of all other existing set points and on their number.

  • comediance
    is similar to covariance. It is a triangular symmetric matrix, obtained by supplying method covar with the geometric median instead of the usual centroid. Thus zero mean vectors are replaced by zero median vectors in the covariance calculations. The results are similar but more stable with respect to the outliers.

  • outer_hull is a subset of all zero median points p, such that no other points lie outside the normal plane through p. The points that do not satisfy this condition are called the internal points.

  • inner_hull is a subset of all zero median points p, that do not lie outside the normal plane of any other point. Note that in a highly dimensional space up to all points may belong to both the inner and the outer hulls, as, for example, when they all lie on the same hypersphere.

  • depth is a measure of likelihood of a zero median point p belonging to a data cloud. More specifically, it is the projection onto unit p of a sum of unit vectors that lie outside the normal through p. For example, all outer hull points have by their definition depth = 0, whereas the inner hull points have high values of depth. This is intended as an improvement on Mahalanobis distance which has a similar goal but says nothing about how well enclosed p is. Whereas tm_statistic only informs about the probability pertaining to the whole cloud, not to its local shape near p.

  • sigvec (signature vector)
    Proportional projections of a cloud of zero median vectors on all hemispheric axis. When a new zero median point p needs to be classified, we can quickly estimate how well populated is its direction from gm. Similar could be done by projecting all the points directly onto p but this is usually impractically slow, as there are typically very many such points. However, signature_vector only needs to be precomputed once and is then the only vector to be projected onto p.

Previously Known Concepts and Terminology

  • centroid/centre/mean of an nd set.
    Is the point, generally non member, that minimises its sum of squares of distances to all member points. The squaring makes it susceptible to outliers. Specifically, it is the d-dimensional arithmetic mean. It is sometimes called 'the centre of mass'. Centroid can also sometimes mean the member of the set which is the nearest to the Centre. Here we follow the common usage: Centroid = Centre = Arithmetic Mean.

  • quasi/marginal median
    is the point minimising sums of distances separately in each dimension (its coordinates are medians along each axis). It is a mistaken concept which we do not recommend using.

  • Tukey median
    is the point maximising Tukey's Depth, which is the minimum number of (outlying) points found in a hemisphere in any direction. Potentially useful concept but its advantages over the geometric median are not clear.

  • true geometric median (gm)
    is the point (generally non member), which minimises the sum of distances to all member points. This is the one we want. It is less susceptible to outliers than the centroid. In addition, unlike quasi median, gm is rotation independent.

  • medoid
    is the member of the set with the least sum of distances to all other members. Equivalently, the member which is the nearest to the gm (has the minimum radius).

  • outlier
    is the member of the set with the greatest sum of distances to all other members. Equivalently, it is the point furthest from the gm (has the maximum radius).

  • Mahalanobis distance
    is a scaled distance, whereby the scaling is derived from the axis of covariances / comediances of the data points cloud. Distances in the directions in which there are few points are increased and distances in the directions of significant covariances / comediances are decreased. Requires matrix decomposition. Mahalanobis distance is defined as: m(d) = sqrt(d'inv(C)d) = sqrt(d'inv(LL')d) = sqrt(d'inv(L')inv(L)d), where inv() denotes matrix inverse, which is never explicitly computed and ' denotes transposition.
    Let x = inv(L)d ( and therefore also x' = d'inv(L') ).
    Substituting x into the above definition: `m(d) = sqrt(x'x) = |x|.
    We obtain x by setting Lx = d and solving by forward substitution.
    All these calculations are done in the compact triangular form.

  • Cholesky-Banachiewicz matrix decomposition
    decomposes any positive definite matrix S (often covariance or comediance matrix) into a product of lower triangular matrix L and its transpose L': S = LL'. The determinant of S can be obtained from the diagonal of L. We implemented the decomposition on TriangMat for maximum efficiency. It is used mainly by mahalanobis.

  • Householder's decomposition
    in cases where the precondition (positive definite matrix S) for the Cholesky-Banachiewicz decomposition is not satisfied, Householder's (UR) decomposition is often used as the next best method. It is implemented here on our efficient struct TriangMat.

  • wedge product, geometric product
    products of the Grassman and Clifford algebras, respectively. Wedge product is used here to generalize the cross product of two vectors into any number of dimensions, determining the correct sign (sidedness of their common plane).

Implementation Notes

The main constituent parts of Rstats are its traits. The different traits are determined by the types of objects to be handled. The objects are mostly vectors of arbitrary length/dimensionality (d). The main traits are implementing methods applicable to:

  • Stats: a single vector (of numbers),
  • Vecg: methods operating on two vectors, e.g. scalar product,
  • Vecu8: some methods specialized for end-type u8,
  • MutVecg: some of the above methods, mutating self,
  • VecVec: methods operating on n vectors (rows of numbers),
  • VecVecg: methods for n vectors, plus another generic argument, typically a vector of n weights, expressing the relative significance of the vectors.

The traits and their methods operate on arguments of their required categories. In classical statistical parlance, the main categories correspond to the number of 'random variables'.

Vec<Vec<T>> type is used for rectangular matrices (could also have irregular rows).

struct TriangMat is used for symmetric / antisymmetric / transposed / triangular matrices and wedge and geometric products. All instances of TriangMat store only n*(n+1)/2 items in a single flat vector, instead of n*n, thus almost halving the memory requirements. Their transposed versions only set up a flag kind >=3 that is interpreted by software, instead of unnecessarily rewriting the whole matrix. Thus saving processing of all transposes (a common operation). All this is put to a good use in our implementation of the matrix decomposition methods.

The vectors' end types (of the actual data) are mostly generic: usually some numeric type. Copy trait bounds on these generic input types have been relaxed to Clone, to allow cloning user's own end data types in any way desired. There is no difference for primitive types.

The computed results end types are usually f64.

Errors

Rstats crate produces custom error RError:

pub enum RError<T> where T:Sized+Debug {
    /// Insufficient data
    NoDataError(T),
    /// Wrong kind/size of data
    DataError(T),
    /// Invalid result, such as prevented division by zero
    ArithError(T),
    /// Other error converted to RError
    OtherError(T)
}

Each of its enum variants also carries a generic payload T. Most commonly this will be a String message, giving more helpful explanation, e.g.:

if dif <= 0_f64 {
    return Err(RError::ArithError("cholesky needs a positive definite matrix".to_owned())));
};

format!(...) can be used to insert (debugging) run-time values to the payload String. These errors are returned and can then be automatically converted (with ?) to users' own errors. Some such error conversions are implemented at the bottom of errors.rs file and used in tests.rs.

There is a type alias shortening return declarations to, e.g.: Result<Vec<f64>,RE>, where

pub type RE = RError<String>;

Convenience functions nodata_error, data_error, arith_error, other_error are used to construct and return these errors. Their message argument can be either literal &str, or String (e.g. constructed by format!). They return ReError<String> already wrapped up as an Err variant of Result. cf.:

if dif <= 0_f64 {
    return arith_error("cholesky needs a positive definite matrix");
};

Structs

struct Params

holds the central tendency of 1d data, e.g. any kind of mean, or median, and any spread measure, e.g. standard deviation or 'mad'.

struct TriangMat

holds triangular matrices of all kinds, as described in Implementation section above. Beyond the expansion to their full matrix forms, a number of (the best) Linear Algebra methods are implemented directly on TriangMat, in module triangmat.rs, such as:

  • Cholesky-Banachiewicz matrix decomposition: S = LL' (where ' denotes the transpose). This decomposition is used by mahalanobis, determinant, etc.

  • Mahalanobis Distance for ML recognition tasks.

  • Various operations on TriangMats, including mult: matrix multiplication of two triangular or symmetric or antisymmetric matrices in this compact form, without their expansions to full matrices.

Also, some methods, specifically the covariance/comedience calculations in VecVec and VecVecg return TriangMat matrices. These are positive definite, which makes the most efficient Cholesky-Banachiewicz decomposition applicable to them.

Similarly, Householder UR (M = QR), which is a more general matrix decomposition, also returns TriangMats.

Quantify Functions (Dependency Injection)

Most methods in medians and some in indxvec crates, e.g. find_any and find_all, require explicit closure passed to them, usually to tell them how to quantify input data of any type T into f64. Variety of different quantifying methods can then be dynamically employed.

For example, in text analysis (&str end type), it can be the word length, or the numerical value of its first few letters, or the numerical value of its consonants, etc. Then we can sort them or find their means / medians / spreads under all these different measures. We do not necessarily want to explicitly store all such different values, as input data can be voluminous. It is often preferable to be able to compute any of them on demand, using these closure arguments.

When data is already of the required end-type, use the 'dummy' closure:

|&f| f

When T is a primitive type, such as i64, u64, usize, that can be converted to f64, possibly with some loss of accuracy, use:

|&f| f as f64

fromop

When T is convertible by an existing custom From implementation (and f64:From<T>, T:Clone have been duly added everywhere as trait bounds), then simply pass in fromop, defined as:

/// Convenience From quantification invocation
pub fn fromop<T: Clone + Into<f64>>(f: &T) -> f64 {
    f.clone().into()
}|

The remaining general cases previously required new manual implementations to be written for the (global) From trait for each new type and for each different quantification method, plus adding its trait bounds everywhere. Even then, the different implementations of From would conflict with each other. Now we can simply implement all the custom quantifications within the closures. This generality is obtained at the price of a small inconvenience: having to supply one of the above closures argument for the primitive types as well.

Auxiliary Functions

  • fromop: see above.

  • sumn: the sum of the sequence 1..n = n*(n+1)/2. It is also the size of a lower/upper triangular matrix.

  • tm_stat: (x-centre)/dispersion. Generalised t-statistic in one dimension.

  • unit_matrix: - generates full square unit matrix.

  • nodata_error, data_error, arith_error, other_error - construct custom RE errors (see section Errors above).

Trait Stats

One dimensional statistical measures implemented for all numeric end types.

Its methods operate on one slice of generic data and take no arguments. For example, s.amean()? returns the arithmetic mean of the data in slice s. These methods are checked and will report RError(s), such as an empty input. This means you have to apply ? to their results to pass the errors up, or explicitly match them to take recovery actions, depending on the error variant.

Included in this trait are:

  • 1d medians (classic, geometric and harmonic) and their spreads
  • 1d means (arithmetic, geometric and harmonic) and their spreads
  • linearly weighted means (useful for time analysis),
  • probability density function (pdf)
  • autocorrelation, entropy
  • linear transformation to [0,1],
  • other measures and basic vector algebra operators

Note that fast implementations of 1d 'classic' medians are, as of version 1.1.0, provided in a separate crate medians.

Trait Vecg

Generic vector algebra operations between two slices &[T], &[U] of any (common) length (dimensions). Note that it may be necessary to invoke some using the 'turbofish' ::<type> syntax to indicate the type U of the supplied argument, e.g.:

datavec.somemethod::<f64>(arg)

Methods implemented by this trait:

  • Vector additions, subtractions and products (scalar, Kronecker, outer),
  • Other relationships and measures of difference,
  • Pearson's, Spearman's and Kendall's correlations,
  • Joint pdf, joint entropy, statistical independence (based on mutual information).
  • Contribution measure of a point's impact on the geometric median

Note that our median correlation is implemented in a separate crate medians.

Some simpler methods of this trait may be unchecked (for speed), so some caution with data is advisable.

Trait MutVecg

A select few of the Stats and Vecg methods (e.g. mutable vector addition, subtraction and multiplication) are reimplemented under this trait, so that they can mutate self in-place. This is more efficient and convenient in some circumstances, such as in vector iterative methods.

However, these methods do not fit in with the functional programming style, as they do not explicitly return anything (their calls are statements with side effects, rather than expressions).

Trait Vecu8

Some vector algebra as above that can be more efficient when the end type happens to be u8 (bytes). These methods have u8 appended to their names to avoid confusion with Vecg methods. These specific algorithms are different to their generic equivalents in Vecg.

  • Frequency count of bytes by their values (histogram, pdf, jointpdf)
  • Entropy, jointentropy, independence.

Trait VecVec

Relationships between n vectors in d dimensions. This (hyper-dimensional) data domain is denoted here as (nd). It is in nd where the main original contribution of this library lies. True geometric median (gm) is found by fast and stable iteration, using improved Weiszfeld's algorithm gmedian. This algorithm solves Weiszfeld's convergence and stability problems in the neighbourhoods of existing set points. Its variant, par_gmedian, employs multithreading for faster execution and gives otherwise the same result.

  • centroid, medoid, outliers, gm
  • sums of distances, radius of a point (as its distance from gm)
  • characterisation of a set of multidimensional points by the mean, standard deviation, median and MAD of its points' radii. These are useful recognition measures for the set.
  • transformation to zero geometric median data,
  • multivariate trend (regression) between two sets of nd points,
  • covariance and comediance matrices.
  • inner and outer hulls

Trait VecVecg

Methods which take an additional generic vector argument, such as a vector of weights for computing weighted geometric medians (where each point has its own significance weight). Matrices multiplications.

Appendix: Recent Releases

  • Version 2.2.12 - Some corrections of Readme.md.

  • Version 2.1.11 - Some minor tidying up of code.

  • Version 2.1.10 - Added project of a TriangMat to a subspace given by a subspace index.

  • Version 2.1.9 - Added multiplications and more tests for TriangMat.

  • Version 2.1.8 - Improved TriangMat::diagonal(), restored TriangMat::determinant(), tidied up triangmat test.

  • Version 2.1.7 - Removed suspect eigen values/vectors computations. Improved 'householder' test.

  • Version 2.1.5 - Added projection to trait VecVecg to project all self vectors to a new basis. This can be used e.g. for Principal Components Analysis data reduction, using some of the eigenvectors as the new basis.

  • Version 2.1.4 - Tidied up some error processing.

  • Version 2.1.3 - Added normalize (normalize columns of a matrix and transpose them to rows).

  • Version 2.1.2 - Added function project to project a TriangMat to a lower dimensional space of selected dimensions. Removed rows which was a duplicate of dim.

  • Version 2.1.0 - Changed the type of mid argument to covariance methods from U -> f64, making the normal expectation for the type of precise geometric medians explicit. Accordingly, moved covar and serial_covar from trait VecVecg to VecVec. This might potentially require changing some use declarations in your code.

  • Version 2.0.12 - added depth_ratio

  • Version 2.0.11 - removed not so useful variances. Tidied up error processing in vecvecg.rs. Added to it serial_covar and serial_wcovar for when heavy loading of all the cores may not be wanted.

  • Version 2.0.9 - Pruned some rarely used methods, simplified gmparts and gmerror, updated dependencies.

  • Version 2.0.8' - Changed initial guess in iterative weighted gm methods to weighted mean. This, being more accurate than plain mean, leads to fewer iterations. Updated some dependencies.

  • Version 2.0.7 - Updated to ran 2.0.

  • Version 2.0.6 - Added convenience method medmad to Stats trait. It packs median and mad into struct Params, similarly to ameanstd and others. Consequently simplified the printouts in some tests.

  • Version 2.0.5 - Corrected wsigvec to also return normalized result. Updated dependency Medians to faster version 3.0.1.

  • Version 2.0.4 - Made a corresponding change: winsideness -> wdepth.

  • Version 2.0.3 - Improved insideness to be projection of a sum of unit vectors instead of just a simple count. Renamed it to depth to avoid confusion. Also some fixes to hulls.

  • Version 2.0.2 - Significantly speeded up insideness and added weighted version winsideness to VecVecg trait.

  • Version 2.0.1 - Added TriangMat::dim() and tidied up some comments.

  • Version 2.0.0 - Renamed MStats -> Params and its variant dispersion -> spread. This may cause some backwards incompatibilities, hence the new major version. Added 'centre' as an argument to dfdt,dvdt,wdvdt, so that it does not have to be recomputed.