RankAggreg {RankAggreg}R Documentation

Weighted Rank Aggregation of partial ordered lists

Description

Performs aggregation of ordered lists based on the ranks (optinally with additional weights) via the Cross-Entropy Monte Carlo algorithm or the Genetic Algorithm.

Usage

RankAggreg(x, k, index.weights = NULL, use.weights = FALSE, method = c("CrossEntropy", "GeneticAlgorithm"), 
distance = c("Spearman", "Kendall"), rho = 0.01, weight = 0.5, N = 5 * k * length(unique(sort(as.vector(x)))), 
error = 0.001, maxIter = 100, popSize = 100, CP = 1, MP = 0.001, informative = FALSE, v1 = NULL, verbose = TRUE)

Arguments

x a matrix of ordered lists to be combined (lists must be in rows)
k size of the top-k list
index.weights a matrix of scores (weights) to be used in the aggregation process. Weights in each row must be ordered either in decreasing or increasing order and must correspond to the elements in x
use.weights boolean, if weights are to be used
method method to be used to perform rank aggregation: Cross Entropy Monte Carlo or Genetic Algorithm (GA)
distance distance to be used which "measures" the similarity of ordered lists
rho (rho*N) is the "quantile" of candidate lists sorted by the function values. Used only by the Cross-Entropy algorithm
weight a learning factor used in the probability update procedure of the algorithm. Used only by the Cross-Entropy algorithm
N a number of samples to be generated by the MCMC; default: 5nk, where n is the number of unique elements in x. Used only by the Cross-Entropy algorithm
error convergence criteria for the Cross-Entropy algorithm
maxIter the maximum iterations allowed; can be used as a stopping criteria for the Genetic Algorithm
popSize population size in each generation of Genetic Algorithm; has no effect if method="CrossEntropy"
CP Cross-over probability for the GA; the default value is 1. It is usually greater than 0.5.
MP Mutation probability. This value should be small and the number of mutations in the population of size popSize and the number of features k is computed as popSize*k*MP. Used only by the GA
informative boolean, if informative is TRUE, use v1 as the initial probability matrix
v1 optional, can be used to specify the initial probability matrix; if v1=NULL, the initial probability matrix is set to 1/n, where n is the number of unique elements in x
verbose boolean, if console output is to be displayed at each iteration

Details

The function performs rank aggregation via the Cross-Entropy Monte Carlo algorithm or the Genetic Algorithm. Both approaches can and should be used when k is relatively large (k > 10). If k is small, one can enumerate all possible candidate lists and find the minimum directly using the BruteAggreg function available in this package.

The Cross-Entropy Monte Carlo algorithm is an iterative procedure for solving difficult combinatorial problems in which it is computationally not feasable to find the solution directly. In the context of rank aggregation, the algorithm searches for the "super"-list which is as close as possible to the ordered lists in x. We use either the Spearman footrule distance or the Kendall's tau to measure the "closeness" of any two ordered lists (or modified by us the weighted versions of these distances). Please refer to the paper in the references for further details.

The Genetic Algorithm requires setting CP and MP parameters which effect the degree of "evolution" in the population. If both CP and MP are small, the algorithms is very conservative and may take a long time to search the solution space of all ordered candidate lists. On the other hand, setting CP and MP (especially MP) large will introduce a large number of mutations in the population which can result in a local optima. Two convergence criteria are used to stop the algorithm. The first being the repetition of the same minimum value of the objective function in five consecutive iterations. If that condition is not met in maxIter number of iterations, the algorithm will stop regardless of the first condition.

Value

top.list Top-k aggregated list
optimal.value the minimum value of the objective function corresponding to the top-k list
sample.size the number of samples generated by the MCMC at each iteration
num.iter the number of iterations until convergence
method which algorithm was used
distance which distance was used

Author(s)

Vasyl Pihur, Somnath Datta, Susmita Datta

References

Pihur, V., Datta, S., and Datta, S. (2007) "Weighted rank aggregation of cluster validation measures: a Monte Carlo cross-entropy approach" Bioinformatics, 23(13):1607-1615

See Also

BruteAggreg

Examples

# rank aggregation without weights
x <- matrix(c("A", "B", "C", "D", "E",
        "B", "D", "A", "E", "C",
        "B", "A", "E", "C", "D",
        "A", "D", "B", "C", "E"), byrow=TRUE, ncol=5)

toplist <- RankAggreg(x, 5, rho=.1)

# weighted rank aggregation
set.seed(100)
w <- matrix(rnorm(20), ncol=5)
w <- t(apply(w, 1, sort))

# using the Cross-Entropy Monte-Carlo algorithm
toplistS <- RankAggreg(x, 5, rho=.1, index.weights=w, use.weights=TRUE)
toplistK <- RankAggreg(x, 5, rho=.1, index.weights=w, use.weights=TRUE, distance="Kendall")

# using the Genetic algorithm
toplistS <- RankAggreg(x, 5, rho=.1, index.weights=w, use.weights=TRUE, method="Genetic")
toplistK <- RankAggreg(x, 5, index.weights=w, use.weights=TRUE, distance="Kendall", method="Genetic")

[Package RankAggreg version 0.1-1 Index]