Introduction

MLearn, the workhorse method of MLInterfaces, has been streamlined to support simpler development.

In 1.*, MLearn included a substantial switch statement, and the external learning function was identified by a string. Many massage tasks were wrapped up in switch case elements devoted to each method. MLearn returned instances of MLOutput, but these had complicated subclasses.

MLearn now takes a signature c(“formula”, “data.frame”, “learnerSchema”, “numeric”), with the expectation that extra parameters captured in ... go to the fitting function. The complexity of dealing with expectations and return values of different machine learning functions is handled primarily by the learnerSchema instances. The basic realizations are that

most learning functions use the formula/data idiom, and additional parameters can go in ...
the problem of converting from the function’s output structures (typically lists, but sometimes also objects with attributes) to the uniform structure delived by MLearn should be handled as generically as possible, but specialization will typically be needed
the conversion process can be handled in most cases using only the native R object returned by the learning function, the data, and the training index set.
some functions, like knn, are so idiosyncratic (lacking formula interface or predict method) that special software is needed to adapt MLearn to work with them

Thus we have defined a learnerSchema class,

library(MLInterfaces) 
library(gbm)
getClass("learnerSchema")

## Class "learnerSchema" [package "MLInterfaces"]
## 
## Slots:
##                                                       
## Name:  packageName   mlFunName   converter   predicter
## Class:   character   character    function    function

along with a constructor used to define a family of schema objects that help MLearn carry out specific tasks of learning.

Some examples

We define interface schema instances with suffix "I".

randomForest has a simple converter:

randomForestI@converter

## function (obj, data, trainInd) 
## {
##     teData = data[-trainInd, ]
##     trData = data[trainInd, ]
##     tepr = predict(obj, teData, type = "response")
##     tesco = predict(obj, teData, type = "prob")
##     trpr = predict(obj, trData, type = "response")
##     trsco = predict(obj, trData, type = "prob")
##     names(tepr) = rownames(teData)
##     names(trpr) = rownames(trData)
##     new("classifierOutput", testPredictions = factor(tepr), testScores = tesco, 
##         trainPredictions = factor(trpr), trainScores = trsco, 
##         RObject = obj)
## }
## <bytecode: 0x55e9cda6ca88>
## <environment: namespace:MLInterfaces>

The job of the converter is to populate as much as the classifierOutput instance as possible. For something like nnet, we can do more:

nnetI@converter

## function (obj, data, trainInd) 
## {
##     teData = data[-trainInd, ]
##     trData = data[trainInd, ]
##     tepr = predict(obj, teData, type = "class")
##     trpr = predict(obj, trData, type = "class")
##     names(tepr) = rownames(teData)
##     names(trpr) = rownames(trData)
##     new("classifierOutput", testPredictions = factor(tepr), testScores = predict(obj, 
##         teData), trainScores = predict(obj, trData), trainPredictions = factor(trpr), 
##         RObject = obj)
## }
## <bytecode: 0x55e9cd5e34f0>
## <environment: namespace:MLInterfaces>

We can get posterior class probabilities.

To obtain the predictions necessary for confusionMatrix computation, we may need the converter to know about parameters used in the fit. Here, closures are used.

knnI(k=3, l=2)@converter

## function (obj, data, trainInd) 
## {
##     kpn = names(obj$traindat)
##     teData = data[-trainInd, kpn]
##     trData = data[trainInd, kpn]
##     tepr = predict(obj, teData, k, l)
##     trpr = predict(obj, trData, k, l)
##     names(tepr) = rownames(teData)
##     names(trpr) = rownames(trData)
##     new("classifierOutput", testPredictions = factor(tepr), testScores = attr(tepr, 
##         "prob"), trainPredictions = factor(trpr), trainScores = attr(trpr, 
##         "prob"), RObject = obj)
## }
## <bytecode: 0x55e9cd845b40>
## <environment: 0x55e9cd84c8b8>

So we can have the following calls:

library(MASS)
data(crabs)
kp = sample(1:200, size=120)
rf1 = MLearn(sp~CL+RW, data=crabs, randomForestI, kp, ntree=100)
rf1

## MLInterfaces classification output container
## The call was:
## MLearn(formula = sp ~ CL + RW, data = crabs, .method = randomForestI, 
##     trainInd = kp, ntree = 100)
## Predicted outcome distribution for test set:
## 
##  B  O 
## 41 39 
## Summary of scores on test set (use testScores() method for details):
##        B        O 
## 0.522375 0.477625

RObject(rf1)

## 
## Call:
##  randomForest(formula = formula, data = trdata, ntree = 100) 
##                Type of random forest: classification
##                      Number of trees: 100
## No. of variables tried at each split: 1
## 
##         OOB estimate of  error rate: 41.67%
## Confusion matrix:
##    B  O class.error
## B 31 26   0.4561404
## O 24 39   0.3809524

knn1 = MLearn(sp~CL+RW, data=crabs, knnI(k=3,l=2), kp)
knn1

## MLInterfaces classification output container
## The call was:
## MLearn(formula = sp ~ CL + RW, data = crabs, .method = knnI(k = 3, 
##     l = 2), trainInd = kp)
## Predicted outcome distribution for test set:
## 
##  B  O 
## 46 32 
## Summary of scores on test set (use testScores() method for details):
##    Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
##  0.5000  0.6667  0.6667  0.7604  1.0000  1.0000

Making new interfaces

A simple example: ada

The ada method of the ada package has a formula interface and a predict method. We can create a learnerSchema on the fly, and then use it:

adaI = makeLearnerSchema("ada", "ada", standardMLIConverter )
arun = MLearn(sp~CL+RW, data=crabs, adaI, kp )
confuMat(arun)

##      predicted
## given  B  O
##     B 36  7
##     O 18 19

RObject(arun)

## Call:
## ada(formula, data = trdata)
## 
## Loss: exponential Method: discrete   Iteration: 50 
## 
## Final Confusion Matrix for Data:
##           Final Prediction
## True value  B  O
##          B 46 11
##          O 17 46
## 
## Train Error: 0.233 
## 
## Out-Of-Bag Error:  0.242  iteration= 6 
## 
## Additional Estimates of number of iterations:
## 
## train.err1 train.kap1 
##         42         42

What is the standardMLIConverter?

standardMLIConverter

## function (obj, data, trainInd) 
## {
##     teData = data[-trainInd, ]
##     trData = data[trainInd, ]
##     tepr = predict(obj, teData)
##     trpr = predict(obj, trData)
##     names(tepr) = rownames(teData)
##     names(trpr) = rownames(trData)
##     new("classifierOutput", testPredictions = factor(tepr), trainPredictions = factor(trpr), 
##         RObject = obj)
## }
## <bytecode: 0x55e9ce257ce0>
## <environment: namespace:MLInterfaces>

Dealing with gbm

The gbm package workhorse fitter is gbm. The formula input must have a numeric response, and the predict method only returns a numeric vector. There is also no namespace. We introduced a gbm2 function

gbm2

## function (formula, data, ...) 
## {
##     requireNamespace("gbm")
##     mf = model.frame(formula, data)
##     resp = model.response(mf)
##     if (!is(resp, "factor")) 
##         stop("dependent variable must be a factor in MLearn")
##     if (length(levels(resp)) != 2) 
##         stop("dependent variable must have two levels")
##     nresp = as.numeric(resp == levels(resp)[2])
##     fwn = formula
##     fwn[[2]] = as.name("nresp")
##     newf = as.formula(fwn)
##     data$nresp = nresp
##     ans = gbm(newf, data = data, ...)
##     class(ans) = "gbm2"
##     ans
## }
## <bytecode: 0x55e9cdf22e10>
## <environment: namespace:MLInterfaces>

that requires a two-level factor response and recodes for use by gbm. It also returns an S3 object of newly defined class gbm2, which only returns a factor. At this stage, we could use a standard interface, but the prediction values will be unpleasant to work with. Furthermore the predict method requires specification of n.trees. So we pass a parameter n.trees.pred.

BgbmI

## function (n.trees.pred = 1000, thresh = 0.5) 
## {
##     makeLearnerSchema("MLInterfaces", "gbm2", MLIConverter.Bgbm(n.trees.pred, 
##         thresh))
## }
## <bytecode: 0x55e9d0a07238>
## <environment: namespace:MLInterfaces>

set.seed(1234)
gbrun = MLearn(sp~CL+RW+FL+CW+BD, data=crabs, BgbmI(n.trees.pred=25000,thresh=.5), kp, n.trees=25000, distribution="bernoulli", verbose=FALSE )
gbrun

## MLInterfaces classification output container
## The call was:
## MLearn(formula = sp ~ CL + RW + FL + CW + BD, data = crabs, .method = BgbmI(n.trees.pred = 25000, 
##     thresh = 0.5), trainInd = kp, n.trees = 25000, distribution = "bernoulli", 
##     verbose = FALSE)
## Predicted outcome distribution for test set:
## 
## FALSE  TRUE 
##    43    37 
## Summary of scores on test set (use testScores() method for details):
##     Min.  1st Qu.   Median     Mean  3rd Qu.     Max. 
## -94.1537  -8.0736  -0.9309  -4.0923   7.0035  69.2158

confuMat(gbrun)

##      predicted
## given FALSE TRUE
##     B    40    3
##     O     3   34

summary(testScores(gbrun))

##     Min.  1st Qu.   Median     Mean  3rd Qu.     Max. 
## -94.1537  -8.0736  -0.9309  -4.0923   7.0035  69.2158

Additional features

The xvalSpec class allows us to specify types of cross-validation, and to control carefully how partitions are formed. More details are provided in the MLprac2_2 vignette.

The MLearn approach to clustering and other forms of unsupervised learning

A learner schema for a clustering method needs to specify clearly the feature distance measure. We will experiment here. Our main requirements are

ExpressionSets are the basic input objects
The typical formula interface would be ~. but one can imagine cases where a factor from phenoData is specified as a ‘response’ to color items, and this will be allowed
a clusteringOutput class will need to be defined to contain the results, and it will propagate the result object from the native learning method.