Лабораторная работа: выбор признаков и дискретизация (CARET, FSelector, arules, Boruta)
Задача 1. Пакет CARET и графический разведочный анализ
if(!require(caret)) install.packages('caret', dependencies = TRUE)
library(caret)Выведем доступные методы модели/методы выбора признаков:
names(getModelInfo())## [1] "ada" "AdaBag" "AdaBoost.M1"
## [4] "adaboost" "amdai" "ANFIS"
## [7] "avNNet" "awnb" "awtan"
## [10] "bag" "bagEarth" "bagEarthGCV"
## [13] "bagFDA" "bagFDAGCV" "bam"
## [16] "bartMachine" "bayesglm" "binda"
## [19] "blackboost" "blasso" "blassoAveraged"
## [22] "bridge" "brnn" "BstLm"
## [25] "bstSm" "bstTree" "C5.0"
## [28] "C5.0Cost" "C5.0Rules" "C5.0Tree"
## [31] "cforest" "chaid" "CSimca"
## [34] "ctree" "ctree2" "cubist"
## [37] "dda" "deepboost" "DENFIS"
## [40] "dnn" "dwdLinear" "dwdPoly"
## [43] "dwdRadial" "earth" "elm"
## [46] "enet" "evtree" "extraTrees"
## [49] "fda" "FH.GBML" "FIR.DM"
## [52] "foba" "FRBCS.CHI" "FRBCS.W"
## [55] "FS.HGD" "gam" "gamboost"
## [58] "gamLoess" "gamSpline" "gaussprLinear"
## [61] "gaussprPoly" "gaussprRadial" "gbm_h2o"
## [64] "gbm" "gcvEarth" "GFS.FR.MOGUL"
## [67] "GFS.LT.RS" "GFS.THRIFT" "glm.nb"
## [70] "glm" "glmboost" "glmnet_h2o"
## [73] "glmnet" "glmStepAIC" "gpls"
## [76] "hda" "hdda" "hdrda"
## [79] "HYFIS" "icr" "J48"
## [82] "JRip" "kernelpls" "kknn"
## [85] "knn" "krlsPoly" "krlsRadial"
## [88] "lars" "lars2" "lasso"
## [91] "lda" "lda2" "leapBackward"
## [94] "leapForward" "leapSeq" "Linda"
## [97] "lm" "lmStepAIC" "LMT"
## [100] "loclda" "logicBag" "LogitBoost"
## [103] "logreg" "lssvmLinear" "lssvmPoly"
## [106] "lssvmRadial" "lvq" "M5"
## [109] "M5Rules" "manb" "mda"
## [112] "Mlda" "mlp" "mlpKerasDecay"
## [115] "mlpKerasDecayCost" "mlpKerasDropout" "mlpKerasDropoutCost"
## [118] "mlpML" "mlpSGD" "mlpWeightDecay"
## [121] "mlpWeightDecayML" "monmlp" "msaenet"
## [124] "multinom" "mxnet" "mxnetAdam"
## [127] "naive_bayes" "nb" "nbDiscrete"
## [130] "nbSearch" "neuralnet" "nnet"
## [133] "nnls" "nodeHarvest" "null"
## [136] "OneR" "ordinalNet" "ordinalRF"
## [139] "ORFlog" "ORFpls" "ORFridge"
## [142] "ORFsvm" "ownn" "pam"
## [145] "parRF" "PART" "partDSA"
## [148] "pcaNNet" "pcr" "pda"
## [151] "pda2" "penalized" "PenalizedLDA"
## [154] "plr" "pls" "plsRglm"
## [157] "polr" "ppr" "pre"
## [160] "PRIM" "protoclass" "qda"
## [163] "QdaCov" "qrf" "qrnn"
## [166] "randomGLM" "ranger" "rbf"
## [169] "rbfDDA" "Rborist" "rda"
## [172] "regLogistic" "relaxo" "rf"
## [175] "rFerns" "RFlda" "rfRules"
## [178] "ridge" "rlda" "rlm"
## [181] "rmda" "rocc" "rotationForest"
## [184] "rotationForestCp" "rpart" "rpart1SE"
## [187] "rpart2" "rpartCost" "rpartScore"
## [190] "rqlasso" "rqnc" "RRF"
## [193] "RRFglobal" "rrlda" "RSimca"
## [196] "rvmLinear" "rvmPoly" "rvmRadial"
## [199] "SBC" "sda" "sdwd"
## [202] "simpls" "SLAVE" "slda"
## [205] "smda" "snn" "sparseLDA"
## [208] "spikeslab" "spls" "stepLDA"
## [211] "stepQDA" "superpc" "svmBoundrangeString"
## [214] "svmExpoString" "svmLinear" "svmLinear2"
## [217] "svmLinear3" "svmLinearWeights" "svmLinearWeights2"
## [220] "svmPoly" "svmRadial" "svmRadialCost"
## [223] "svmRadialSigma" "svmRadialWeights" "svmSpectrumString"
## [226] "tan" "tanSearch" "treebag"
## [229] "vbmpRadial" "vglmAdjCat" "vglmContRatio"
## [232] "vglmCumulative" "widekernelpls" "WM"
## [235] "wsrf" "xgbDART" "xgbLinear"
## [238] "xgbTree" "xyf"Создадим тестовый набор данных и выполним графический разведочный
анализ с помощью featurePlot() из пакета
caret.
x <- matrix(rnorm(50*5), ncol = 5)
colnames(x) <- paste0('V', 1:5)
y <- factor(rep(c('A','B'), each = 25))
df <- data.frame(x, Class = y)
# Пары
featurePlot(x = x, y = y, plot = 'pairs')
# Boxplots
featurePlot(x = x, y = y, plot = 'box')
# Плотности
featurePlot(x = x, y = y, plot = 'density')
Выводы:
Графики помогают визуально оценить различия признаков между классами. Признаки с явными различиями распределений полезны для классификации.
Задача 2. FSelector (iris)
if(!require(FSelector)) install.packages('FSelector')
library(FSelector)
data(iris)Рассчитаем несколько метрик важности:
ig <- information.gain(Species ~ ., iris)
gr <- gain.ratio(Species ~ ., iris)
chi <- chi.squared(Species ~ ., iris)
res <- data.frame(InformationGain = ig$attr_importance,
GainRatio = gr$attr_importance,
ChiSquared = chi$attr_importance,
row.names = rownames(ig))
res## InformationGain GainRatio ChiSquared
## Sepal.Length 0.4521286 0.4196464 0.6288067
## Sepal.Width 0.2672750 0.2472972 0.4922162
## Petal.Length 0.9402853 0.8584937 0.9346311
## Petal.Width 0.9554360 0.8713692 0.9432359barplot(ig$attr_importance[order(ig$attr_importance, decreasing = TRUE)],
main = 'Information Gain (iris)', las = 2)
Выводы:
Petal.Length и Petal.Width являются
наиболее информативными признаками для определения
Species.
Задача 3. Дискретизация с помощью arules::discretize() (iris)
if(!require(arules)) install.packages('arules')
library(arules)
vec <- iris$Sepal.Length
# interval: равная ширина интервалов
int_cut <- discretize(vec, method = 'interval', categories = 3)
# frequency: равная частота
freq_cut <- discretize(vec, method = 'frequency', categories = 3)
# cluster: кластеризация (kmeans)
clust_cut <- discretize(vec, method = 'cluster', categories = 3)
# fixed: задаём границы (пример)
fixed_cut <- discretize(vec, method = 'fixed', categories = c(4.3, 5.8))
table(int_cut)## int_cut
## [4.3,5.5) [5.5,6.7) [6.7,7.9]
## 52 70 28table(freq_cut)## freq_cut
## [4.3,5.4) [5.4,6.3) [6.3,7.9]
## 46 53 51table(clust_cut)## clust_cut
## [4.3,5.42) [5.42,6.39) [6.39,7.9]
## 52 56 42table(fixed_cut)## fixed_cut
## [4.3,5.8]
## 80plot(int_cut, main='interval')
plot(freq_cut, main='frequency')
plot(clust_cut, main='cluster')
plot(fixed_cut, main='fixed')
Выводы:
interval может дать неравномерные по численности
категории; frequency — равные по численности;
cluster — полезно при выраженной кластерной структуре;
fixed — для семантически интерпретируемых границ.
Задача 4. Boruta и набор Ozone (или airquality)
if(!require(Boruta)) install.packages('Boruta')
library(Boruta)
# Убедимся, что каталог plots существует
if(!dir.exists('plots')) dir.create('plots')
# Попытка загрузить data('Ozone'); если нет — используем airquality
data_found <- tryCatch({
data(Ozone)
exists('Ozone')
}, error = function(e) FALSE)
if(!data_found){
message('data("Ozone") не найдена — используем airquality')
data_df <- na.omit(airquality)
target <- 'Ozone'
} else {
data_df <- na.omit(get('Ozone'))
target <- names(data_df)[1]
}
# Бинаризуем целевую переменную по медиане для задачи классификации
median_val <- median(data_df[[target]], na.rm = TRUE)
data_df$TargetBin <- factor(ifelse(data_df[[target]] > median_val, 'High', 'Low'))
boruta_formula <- as.formula('TargetBin ~ .')
# Убираем исходную числовую целевую колонку, если она присутствует
boruta_data <- data_df[, !(names(data_df) %in% c(target))]
# Запуск Boruta с установкой случайного семени для воспроизводимости
set.seed(123)
bor <- Boruta(boruta_formula, data = boruta_data, doTrace = 0, maxRuns = 100)
# Получение статистики признаков
final <- TentativeRoughFix(bor)
print(attStats(final))## meanImp medianImp minImp maxImp normHits decision
## Solar.R 10.606715 10.549115 9.140416 12.059908 1 Confirmed
## Wind 10.357017 10.428487 9.040677 12.175220 1 Confirmed
## Temp 33.300052 32.838303 31.929518 35.973780 1 Confirmed
## Month 6.087695 5.661576 4.933978 8.146074 1 Confirmed
## Day 4.582321 4.291928 3.281392 5.742789 1 Confirmedplot(bor, las = 2, cex.axis = 0.7)
Выводы:
Boruta даёт основанные на случайном лесе оценки важности и позволяет определить подтверждённые и отклонённые признаки.