Raw File
Tip revision: d51529a8ee2827cd8c1bd7c4a9e0265dd0cc72cf authored by Toni Giorgino on 19 September 2022, 16:36:11 UTC
version 1.23-1
Tip revision: d51529a
% Generated by roxygen2: do not edit by hand
% Please edit documentation in R/dtwDist.R
\title{Compute a dissimilarity matrix}
dtwDist(mx, my = mx, ...)
\item{mx}{numeric matrix, containing timeseries as rows}

\item{my}{numeric matrix, containing timeseries as rows (for cross-distance)}

\item{...}{arguments passed to the \code{\link[=dtw]{dtw()}} call}
A square matrix whose element \verb{[i,j]} holds the Dynamic Time
Warp distance between row \code{i} (query) and \code{j} (reference) of
\code{mx} and \code{my}, i.e.  \code{dtw(mx[i,],my[j,])$distance}.
Compute the dissimilarity matrix between a set of single-variate timeseries.
\code{dtwDist} computes a dissimilarity matrix, akin to \code{\link[=dist]{dist()}},
based on the Dynamic Time Warping definition of a distance between
single-variate timeseries.

The \code{dtwDist} command is a synonym for the \code{\link[proxy:dist]{proxy::dist()}}
function of package \CRANpkg{proxy}; the DTW distance is registered as
\code{method="DTW"} (see examples below).

The timeseries are stored as rows in the matrix argument \code{m}. In other
words, if \code{m} is an N * T matrix, \code{dtwDist} will build N\emph{N ordered
pairs of timeseries, perform the corresponding N}N \code{dtw} alignments,
and return all of the results in a matrix. Each of the timeseries is T
elements long.

\code{dtwDist} returns a square matrix, whereas the \code{dist} object is
lower-triangular. This makes sense because in general the DTW "distance" is
not symmetric (see e.g.  asymmetric step patterns).  To make a square matrix
with the \code{\link[proxy:dist]{proxy::dist()}} function semantics, use the two-arguments
call as \code{dist(m,m)}. This will return a square \code{crossdist} object.
To convert a square cross-distance matrix (\code{crossdist} object) to
a symmetric \code{\link[=dist]{dist()}} object, use a suitable conversion strategy
(see examples).

## Symmetric step pattern => symmetric dissimilarity matrix;
## no problem coercing it to a dist object:

m <- matrix(0,ncol=3,nrow=4)
m <- row(m)

# Old-fashioned call style would be:
#   dtwDist(m)
#   as.dist(dtwDist(m))

## Find the optimal warping _and_ scale factor at the same time.
## (There may be a better, analytic way)

# Prepare a query and a reference


# Make a set of several references, scaled from 0 to 3 in .1 increments.
# Put them in a matrix, in rows

scaleSet <- seq(0.1,3,by=.1)

# The query has to be made into a 1-row matrix.
# Perform all of the alignments at once, and normalize the result.


# The optimal scale for the reference is 1.0
  xlab="Reference scale factor (denominator)",
  ylab="DTW distance",type="o",
  main="Sine vs scaled cosine alignment, 0 to 4 pi")

## Asymmetric step pattern: we can either disregard part of the pairs
## (as.dist), or average with the transpose

mm <- matrix(runif(12),ncol=3)
dm <- dist(mm,mm,method="DTW",step=asymmetric); # a crossdist object

# Old-fashioned call style would be:
#   dm <- dtwDist(mm,step=asymmetric)
#   as.dist(dm)

## Symmetrize by averaging:

## check definition

Toni Giorgino
back to top