単純ベイズ分類器を使用したSMSスパムのフィルタリング（Rのコード）

 PreprocessSentence <- function(s) { # Cut and make some preprocessing with input sentence words <- strsplit(gsub(pattern="[[:digit:]]+", replacement="1", x=tolower(s)), '[[:punct:][:blank:]]+') return(words) } LoadData <- function(fileName = "./Data/Spam/SMSSpamCollection") { # Read data from text file and makes simple preprocessing: # to lower case -> replace all digit strings with 1 -> split with punctuation and blank characters con <- file(fileName,"rt") lines <- readLines(con) close(con) df <- data.frame(lab = rep(NA, length(lines)), data = rep(NA, length(lines))) for(i in 1:length(lines)) { tmp <- unlist(strsplit(lines[i], '\t', fixed = T)) df$lab[i] <- tmp[1] df$data[i] <- PreprocessSentence(tmp[2]) } return(df) } 

 CreateDataSet <- function(dataSet, proportions = c(0.6, 0.2, 0.2)) { # Creates a list with indices of train, validation and test sets proportions <- proportions/sum(proportions) hamIdx <- which(df$lab == "ham") nham <- length(hamIdx) spamIdx <- which(df$lab == "spam") nspam <- length(spamIdx) hamTrainIdx <- sample(hamIdx, floor(proportions[1]*nham)) hamIdx <- setdiff(hamIdx, hamTrainIdx) spamTrainIdx <- sample(spamIdx, floor(proportions[1]*nspam)) spamIdx <- setdiff(spamIdx, spamTrainIdx) hamValidationIdx <- sample(hamIdx, floor(proportions[2]*nham)) hamIdx <- setdiff(hamIdx, hamValidationIdx) spamValidationIdx <- sample(spamIdx, floor(proportions[2]*nspam)) spamIdx <- setdiff(spamIdx, spamValidationIdx) ds <- list( train = sample(union(hamTrainIdx, spamTrainIdx)), validation = sample(union(hamValidationIdx, spamValidationIdx)), test = sample(union(hamIdx, spamIdx)) ) return(ds) } 

 CreateModel <- function(data, laplaceFactor = 0) { # creates naive bayes spam classifier based on data m <- list(laplaceFactor = laplaceFactor) m[["total"]] <- length(data$lab) m[["ham"]] <- list() m[["spam"]] <- list() m[["hamLabelCount"]] <- sum(data$lab == "ham") m[["spamLabelCount"]] <- sum(data$lab == "spam") m[["hamWordCount"]] <- 0 m[["spamWordCount"]] <- 0 uniqueWordSet <- c() for(i in 1:length(data$lab)) { sentence <- unlist(data$data[i]) uniqueWordSet <- union(uniqueWordSet, sentence) for(j in 1:length(sentence)) { if(data$lab[i] == "ham") { if(is.null(m$ham[[sentence[j]]])) { m$ham[[sentence[j]]] <- 1 } else { m$ham[[sentence[j]]] <- m$ham[[sentence[j]]] + 1 } m[["hamWordCount"]] <- m[["hamWordCount"]] + 1 } else if(data$lab[i] == "spam") { if(is.null(m$spam[[sentence[j]]])) { m$spam[[sentence[j]]] <- 1 } else { m$spam[[sentence[j]]] <- m$spam[[sentence[j]]] + 1 } m[["spamWordCount"]] <- m[["spamWordCount"]] + 1 } } } m[["uniqueWordCount"]] <- length(uniqueWordSet) return(m) } 

 ClassifySentense <- function(s, model, preprocess = T) { # calculate class of the input sentence based on the model GetCount <- function(w, ls) { if(is.null(ls[[w]])) { return(0) } return(ls[[w]]) } words <- unlist(s) if(preprocess) { words <- unlist(PreprocessSentence(s)) } ham <- log(model$hamLabelCount/(model$hamLabelCount + model$spamLabelCount)) spam <- log(model$spamLabelCount/(model$hamLabelCount + model$spamLabelCount)) for(i in 1:length(words)) { ham <- ham + log((GetCount(words[i], model$ham) + model$laplaceFactor) /(model$hamWordCount + model$laplaceFactor*model$uniqueWordCount)) spam <- spam + log((GetCount(words[i], model$spam) + model$laplaceFactor) /(model$spamWordCount + model$laplaceFactor*model$uniqueWordCount)) } if(ham >= spam) { return("ham") } return("spam") } 

 TestModel <- function(data, model) { # calculate percentage of errors errors <- 0 for(i in 1:length(data$lab)) { predictedLabel <- ClassifySentense(data$data[i], model, preprocess = F) if(predictedLabel != data$lab[i]) { errors <- errors + 1 } } return(errors/length(data$lab)) } 

 CrossValidation <- function(trainData, validationData, laplaceFactorValues, showLog = F) { cvErrors <- rep(NA, length(laplaceFactorValues)) for(i in 1:length(laplaceFactorValues)) { model <- CreateModel(trainData, laplaceFactorValues[i]) cvErrors[i] <- TestModel(validationData, model) if(showLog) { print(paste(laplaceFactorValues[i], ": error is ", cvErrors[i], sep="")) } } return(cvErrors) }