mi_learn: Multiple Instance Learning

In kuwisdelu/matter: Scientific computing for out-of-memory signals and images

Multiple Instance Learning

Description

Multiple instance learning is a strategy for training classifiers when the class labels are observed at a coarser level than the individual data points. For example, if an entire image is classified as "positive" or "negative" but the classifier is trained and predicts at the pixel level.

Usage

mi_learn(fn, x, y, bags, pos = 1L, ...,
	score = fitted, threshold = 0.01, verbose = NA)

Arguments

fn	The function used to train the classifier.
x	The data matrix.
y	The response. This can be the same length as the data, or the same length as the number of bags. Must have exactly two levels when coerced to a factor.
bags	The bags to which the data points belong. The class labels are observed per-bag rather than per data point.
pos	The positive class label, as a string matching one of the levels of y, or the index of the level.
...	Additional options passed to fn.
score	The function used to extract the scores for prediction.
threshold	The stopping criterion. The learning stops when the proportion of updated labels between iterations is less than this value.
verbose	Should progress be printed for each iteration? Not passed to fn.

Details

This is a generic wrapper for applying a multiple instance learning strategy for any classifier that satisfies certain criteria. The labels must be binary (positive and negative).

The multiple instance learning algorithm here assumes that if a single data point is positive, then the entire bag to which it belongs is labeled as positive. For example, if a single pixel in an image indicates the presence of disease, then the entire image is labeled as disease.

The model returned by fn must support returning either a vector of probabilities or a 2-column score matrix when passed to score.

Value

A model object returned by fn.

Author(s)

Kylie A. Bemis

References

D. Guo, M. C. Foell, V. Volkmann, K. Enderle-Ammour, P. Bronsert, O. Shilling, and O. Vitek. “Deep multiple instance learning classifies subtissue locations in mass spectrometry images from tissue-level annotations.” Bioinformatics, vol. 36, issue Supplement_1, pp. i300-i308, 2020.

See Also

nscentroids

Examples

register(SerialParam())
set.seed(1)
n <- 100
p <- 5
g <- 5
bags <- rep(paste0("s", seq_len(g)), each=n %/% g)
bags <- factor(rep_len(bags, n))
x <- matrix(rnorm(n * p), nrow=n, ncol=p)
colnames(x) <- paste0("x", seq_len(p))

# create bagged labels
y <- ifelse(bags %in% c("s1", "s2", "s3"), "pos", "neg")
y <- factor(y, levels=c("pos", "neg"))
ipos <- which(y == "pos")
ineg <- which(y == "neg")
z <- y

# create "true" labels (with some within-bag noise)
z[ipos] <- sample(c("pos", "neg"), length(ipos), replace=TRUE)
jpos <- which(z == "pos")
jneg <- which(z == "neg")

# create data
x[jpos,] <- x[jpos,] + rnorm(p * length(jpos), mean=1)
x[jneg,] <- x[jneg,] - rnorm(p * length(jneg), mean=1)

# fit ordinary NSC and mi-NSC
fit0 <- nscentroids(x=x, y=y)
fit1 <- mi_learn(nscentroids, x=x, y=y, bags=bags, priors=1)

# improved performance on "true" labels
mean(fitted(fit0, "class") == z)
mean(fitted(fit1, "class") == z)

kuwisdelu/matter documentation built on May 1, 2024, 5:17 a.m.