| Title: | Explore Classification Models in High Dimensions |
|---|---|
| Description: | Given $p$-dimensional training data containing $d$ groups (the design space), a classification algorithm (classifier) predicts which group new data belongs to. Generally the input to these algorithms is high dimensional, and the boundaries between groups will be high dimensional and perhaps curvilinear or multi-faceted. This package implements methods for understanding the division of space between the groups. |
| Authors: | Hadley Wickham <[email protected]> |
| Maintainer: | Hadley Wickham <[email protected]> |
| License: | MIT + file LICENSE |
| Version: | 0.4.1.9000 |
| Built: | 2024-10-05 02:39:24 UTC |
| Source: | https://github.com/hadley/classifly |
This is used to identify the boundaries between classification regions. Points with low (close to 0) advantage are likely to be near boundaries.
advantage(post)advantage(post)
post |
matrix of posterior probabilities |
This is a convenient function to fit a classification function and
then explore the results using GGobi. You can also do this in two
separate steps using the classification function and then
explore.
classifly( data, model, classifier, ..., n = 10000, method = "nonaligned", type = "range" )classifly( data, model, classifier, ..., n = 10000, method = "nonaligned", type = "range" )
data |
Data set use for classification |
model |
Classification formula, usually of the form
|
classifier |
Function to use for the classification, eg.
|
... |
Other arguments passed to classification function. For
example. if you use |
n |
Number of points to simulate. To maintain the illusion of a filled solid this needs to increase with dimension. 10,000 points seems adequate for up to four of five dimensions, but if you have more predictors than that, you will need to increase this number. |
method |
method to simulate points: grid, random or nonaligned
(default). See |
type |
type of scaling to apply to data. Defaults to commmon range.
See |
By default in GGobi, points that are not on the boundary (ie. that have an advantage greater than the 5 to brush mode and choose include shadowed points from the brush menu on the plot window. You can then brush them yourself to explore how the certainty of classification varies throughout the space
Special notes:
You should make sure the response variable is a factor
For SVM, make sure to include probability = TRUE in the
arguments to classifly
explore, http://had.co.nz/classifly
data(kyphosis, package = "rpart") library(MASS) classifly(kyphosis, Kyphosis ~ . , lda) classifly(kyphosis, Kyphosis ~ . , qda) classifly(kyphosis, Kyphosis ~ . , glm, family="binomial") classifly(kyphosis, Kyphosis ~ . , knnf, k=3) library(rpart) classifly(kyphosis, Kyphosis ~ . , rpart) if (require("e1071")) { classifly(kyphosis, Kyphosis ~ . , svm, probability=TRUE) classifly(kyphosis, Kyphosis ~ . , svm, probability=TRUE, kernel="linear") classifly(kyphosis, Kyphosis ~ . , best.svm, probability=TRUE, kernel="linear") # Also can use explore directorly bsvm <- best.svm(Species~., data = iris, gamma = 2^(-1:1), cost = 2^(2:+ 4), probability=TRUE) explore(bsvm, iris) }data(kyphosis, package = "rpart") library(MASS) classifly(kyphosis, Kyphosis ~ . , lda) classifly(kyphosis, Kyphosis ~ . , qda) classifly(kyphosis, Kyphosis ~ . , glm, family="binomial") classifly(kyphosis, Kyphosis ~ . , knnf, k=3) library(rpart) classifly(kyphosis, Kyphosis ~ . , rpart) if (require("e1071")) { classifly(kyphosis, Kyphosis ~ . , svm, probability=TRUE) classifly(kyphosis, Kyphosis ~ . , svm, probability=TRUE, kernel="linear") classifly(kyphosis, Kyphosis ~ . , best.svm, probability=TRUE, kernel="linear") # Also can use explore directorly bsvm <- best.svm(Species~., data = iris, gamma = 2^(-1:1), cost = 2^(2:+ 4), probability=TRUE) explore(bsvm, iris) }
The default method currently works for classification functions.
explore(model, data, n = 10000, method = "nonaligned", advantage = TRUE, ...)explore(model, data, n = 10000, method = "nonaligned", advantage = TRUE, ...)
model |
classification object |
data |
data set used with classifier |
n |
number of points to generate when searching for boundaries |
method |
method to generate points, see |
advantage |
only display boundaries |
... |
other arguments not currently used |
It generates a data set filling the design space, finds class boundaries (if desired) and then displays in a new ggobi instance.
A invisible data frame of class classifly
that contains all the simulated and true data. This can be saved and
then printed later to open with rggobi.
generate_classification_data,
http://had.co.nz/classifly
if (require("e1071")) { bsvm <- best.svm(Species~., data = iris, gamma = 2^(-1:1), cost = 2^(2:+ 4), probability=TRUE) explore(bsvm, iris) }if (require("e1071")) { bsvm <- best.svm(Species~., data = iris, gamma = 2^(-1:1), cost = 2^(2:+ 4), probability=TRUE) explore(bsvm, iris) }
Given a model, this function generates points within the range of the data, classifies them, and attempts to locate boundaries by looking at advantage.
generate_classification_data(model, data, n, method, advantage)generate_classification_data(model, data, n, method, advantage)
model |
classification model |
data |
data set used in model |
n |
number of points to generate |
method |
method to use, currently either grid (an evenly spaced grid), random (uniform random distribution across cube), or nonaligned (grid + some random peturbationb) |
advantage |
if |
If posterior probabilities of classification are available, then the
advantage will be calculated directly. If not,
knn is used calculate the advantage based on the number of
neighbouring points that share the same classification. Because knn is
$O(n^2)$ this method is rather slow for large (>20,000 say) data sets.
By default, the boundary points are identified as those below the 5th-percentile for advantage.
data.frame of classified data
This method generates new data that fills the range of the supplied datasets.
generate_data(data, n = 10000, method = "grid")generate_data(data, n = 10000, method = "grid")
data |
data frame |
n |
desired number of new observations |
method |
method to use, see |
knn to allow use
with classifly.A wrapper function for knn to allow use
with classifly.
knnf(formula, data, k = 2)knnf(formula, data, k = 2)
formula |
classification formula |
data |
training data set |
k |
number of neighbours to use |
The olive oil data consists of the percentage composition of 8 fatty acids (palmitic, palmitoleic, stearic, oleic, linoleic, linolenic, arachidic, eicosenoic) found in the lipid fraction of 572 Italian olive oils. There are 9 collection areas, 4 from southern Italy (North and South Apulia, Calabria, Sicily), two from Sardinia (Inland and Coastal) and 3 from northern Italy (Umbria, East and West Liguria).
A data frame with 244 rows and 7 variables
Forina, M. and Armanino, C. and Lanteri, S. and Tiscornia, E., Classification of olive oils from their fatty acid composition, 1983, in Food Research and Data Analysis, edited by Martens, H. and Russwurm Jr, H, pages 189-214.
Every classification method seems to provide a slighly different way of retrieving the posterior probability of group membership. This function provides a common interface to all of them
posterior(model, data)posterior(model, data)
model |
model object |
data |
data set used in model |
Given a vector of data this function will simulate data that could have come from that vector.
simvar(x, n = 10, method = "grid")simvar(x, n = 10, method = "grid")
x |
data vector |
n |
desired number of points (will not always be achieved) |
method |
grid simulation method. See details. |
There are three methods to choose from:
nonaligned (default): grid + some random peturbation
grid: grid of evenly spaced observations. If a factor, all levels in a factor will be used, regardless of n
random: a random uniform sample from the range of the variable
Due to the way that most model objects are stored, you also need to supply the data set you used with the original data set. It currently doesn't support models fitted without using a data argument.
variables(model)variables(model)
model |
model object |
list containing response and predictor variables