kernlab (version 0.9-13)

ranking: Ranking

Description

A universal ranking algorithm which assigns importance/ranking to data points given a query.

Usage

## S3 method for class 'matrix':
ranking(x, y,  kernel ="rbfdot", kpar = list(sigma = 1),  
        scale = FALSE, alpha = 0.99, iterations = 600, 
        edgegraph = FALSE, convergence = FALSE ,...)
## S3 method for class 'kernelMatrix':
ranking(x, y, alpha = 0.99, iterations = 600,
                convergence = FALSE,...)
## S3 method for class 'list':
ranking(x, y, kernel = "stringdot", kpar =
                 list(length = 4, lambda = 0.5), alpha = 0.99, iterations = 600, convergence = FALSE, ...)

Arguments

x

a matrix containing the data to be ranked, or the kernel matrix of data to be ranked or a list of character vectors

y

The index of the query point in the data matrix or a vector of length equal to the rows of the data matrix having a one at the index of the query points index and zero at all the other points.

kernel

the kernel function used in training and predicting. This parameter can be set to any function, of class kernel, which computes a dot product between two vector arguments. kernlab provides the most popular kernel functions which can be used by

kpar

the list of hyper-parameters (kernel parameters). This is a list which contains the parameters to be used with the kernel function. For valid parameters for existing kernels are :

• sigmainverse kernel width for the Radial B

scale

If TRUE the data matrix columns are scaled to zero mean and unit variance.

alpha

The alpha parameter takes values between 0 and 1 and is used to control the authoritative scores received from the unlabeled points. For 0 no global structure is found the algorithm ranks the points similarly to the original dista

iterations

Maximum number of iterations

edgegraph

Construct edgegraph (only supported with the RBF kernel)

convergence

Include convergence matrix in results

...

Additional arguments

Value

• An S4 object of class ranking which extends the matrix class. The first column of the returned matrix contains the original index of the points in the data matrix the second column contains the final score received by each point and the third column the ranking of the point. The object contains the following slots :
• edgegraph
  Containing the edgegraph of the data points.
• convergence
  Containing the convergence matrix

Details

A simple universal ranking algorithm which exploits the intrinsic global geometric structure of the data. In many real world applications this should be superior to a local method in which the data are simply ranked by pairwise Euclidean distances. Firstly a weighted network is defined on the data and an authoritative score is assigned to each query. The query points act as source nodes that continually pump their authoritative scores to the remaining points via the weighted network and the remaining points further spread the scores they received to their neighbors. This spreading process is repeated until convergence and the points are ranked according to their score at the end of the iterations.

References

D. Zhou, J. Weston, A. Gretton, O. Bousquet, B. Schoelkopf Ranking on Data Manifolds Advances in Neural Information Processing Systems 16. MIT Press Cambridge Mass. 2004 http://www.kyb.mpg.de/publications/pdfs/pdf2334.pdf

See Also

ranking-class, specc

Examples

data(spirals)

## create data from spirals
ran <- spirals[rowSums(abs(spirals) < 0.55) == 2,]

## rank points according to similarity to the most upper left point  
ranked <- ranking(ran, 54, kernel = "rbfdot", kpar = list(sigma = 100), edgegraph = TRUE)
ranked[54, 2] <- max(ranked[-54, 2])
c<-1:86
op <- par(mfrow = c(1, 2),pty="s")
plot(ran)
plot(ran, cex=c[ranked[,3]]/40)