dca: Discriminative Component Analysis

In terrytangyuan/dml: Distance Metric Learning in R

Discriminative Component Analysis

Description

Performs discriminative component analysis on the given data.

Usage

dca(data, chunks, neglinks, useD = NULL)

Arguments

data	n * d data matrix. n is the number of data points, d is the dimension of the data. Each data point is a row in the matrix.
chunks	length n vector describing the chunklets: -1 in the i th place means point i doesn't belong to any chunklet; integer j in place i means point i belongs to chunklet j. The chunklets indexes should be 1:(number of chunklets).
neglinks	s * s symmetric matrix describing the negative relationship between all the s chunklets. For the element neglinks_{ij}: neglinks_{ij} = 1 means chunklet i and chunklet j have negative constraint(s); neglinks_{ij} = 0 means chunklet i and chunklet j don't have negative constraints or we don't have information about that.
useD	Integer. Optional. When not given, DCA is done in the original dimension and B is full rank. When useD is given, DCA is preceded by constraints based LDA which reduces the dimension to useD. B in this case is of rank useD.

Value

list of the DCA results:

B	DCA suggested Mahalanobis matrix
DCA	DCA suggested transformation of the data. The dimension is (original data dimension) * (useD)
newData	DCA transformed data

For every two original data points (x1, x2) in newData (y1, y2):

(x2 - x1)' * B * (x2 - x1) = || (x2 - x1) * A ||^2 = || y2 - y1 ||^2

References

Steven C.H. Hoi, W. Liu, M.R. Lyu and W.Y. Ma (2006). Learning Distance Metrics with Contextual Constraints for Image Retrieval. Proceedings IEEE Conference on Computer Vision and Pattern Recognition (CVPR2006).

Examples

## Not run: 
# generate synthetic Gaussian data
library("MASS")

k <- 100 # sample size of each class
n <- 3 # specify how many class
N <- k * n # total sample number

set.seed(123)
x1 <- mvrnorm(k, mu = c(-10, 6), matrix(c(10, 4, 4, 10), ncol = 2))
x2 <- mvrnorm(k, mu = c(0, 0), matrix(c(10, 4, 4, 10), ncol = 2))
x3 <- mvrnorm(k, mu = c(10, -6), matrix(c(10, 4, 4, 10), ncol = 2))
data <- as.data.frame(rbind(x1, x2, x3))

# The fully labeled data set with 3 classes
plot(data$V1, data$V2,
     bg = c("#E41A1C", "#377EB8", "#4DAF4A")[gl(n, k)],
     pch = c(rep(22, k), rep(21, k), rep(25, k))
)
Sys.sleep(3)

# The same data unlabeled; clearly the class structure is less evident
plot(data$V1, data$V2)
Sys.sleep(3)

chunk1 <- sample(1:100, 5)
chunk2 <- sample(setdiff(1:100, chunk1), 5)
chunk3 <- sample(101:200, 5)
chunk4 <- sample(setdiff(101:200, chunk3), 5)
chunk5 <- sample(201:300, 5)
chks <- list(chunk1, chunk2, chunk3, chunk4, chunk5)
chunks <- rep(-1, 300)

# positive samples in the chunks
for (i in 1:5) {
  for (j in chks[[i]]) {
    chunks[j] <- i
  }
}

# define the negative constrains between chunks
neglinks <- matrix(c(
  0, 0, 1, 1, 1,
  0, 0, 1, 1, 1,
  1, 1, 0, 0, 0,
  1, 1, 0, 0, 1,
  1, 1, 1, 1, 0
),
ncol = 5, byrow = TRUE
)

dcaData <- dca(data = data, chunks = chunks, neglinks = neglinks)$newData

# plot DCA transformed data
plot(dcaData[, 1], dcaData[, 2],
     bg = c("#E41A1C", "#377EB8", "#4DAF4A")[gl(n, k)],
     pch = c(rep(22, k), rep(21, k), rep(25, k)),
     xlim = c(-15, 15), ylim = c(-15, 15)
)
## End(Not run)

terrytangyuan/dml documentation built on July 19, 2023, 2 p.m.