从零开始的 k-NN 算法精度低

Question

我正在尝试实施 k-NN 算法，但它总是导致非常低的准确度值。一定有逻辑错误，但我不知道它在哪里。代码如下：

start <- Sys.time()

AccEuc <- NULL
AccMan <- NULL

for(K in grid){
    
    cvAccEuc <- NULL
    cvAccMan <- NULL
    
    for (fold in 1:10){
        
        split = kfoldsplit(dfmerged,10,fold)
        train <- split[[1]][,-c(1,2)]
        valid <- split[[2]][,-c(1,2)]
        trainclass <- split[[1]][,2]
        validclass <- split[[2]][,2]
        
        combined=rbind(valid,train)
        eucdistcombined = as.matrix(dist(combined, method = "euclidian")) # Euclidian Distance
        mandistcombined = as.matrix(dist(combined, method = "manhattan")) # Manhattan Distance
        
        lnvalid = dim(valid)[1]; lntrain = dim(train)[1]; lnall = lnvalid + lntrain
        
        eucdistcombined = eucdistcombined[1:lnvalid,(lnvalid+1):lnall]
        mandistcombined = mandistcombined[1:lnvalid,(lnvalid+1):lnall]
        
        neighbors_euc = t(apply(eucdistcombined, 1, order))
        neighbors_man = t(apply(mandistcombined, 1, order))
        
        idxeuc = t(apply(neighbors_euc, 1, function(x)( x <= K) ))
        idxman = t(apply(neighbors_man, 1, function(x)( x <= K) ))
        
        predseuc <- apply(idxeuc, 1, function(x) as.numeric(getmode( trainclass[as.vector(x)] )) )
        predsman <- apply(idxman, 1, function(x) as.numeric(getmode( trainclass[as.vector(x)] )) )
                          
        cvAccEuc <- c( cvAccEuc, sum(validclass == predseuc)/dim(validclass)[1] )
        cvAccMan <- c( cvAccMan, sum(validclass == predsman)/dim(validclass)[1] )
    }
    
    AccEuc <- rbind(AccEuc, t(c(K,cvAccEuc)))
    AccMan <- rbind(AccMan, t(c(K,cvAccMan)))
}
Sys.time() - start

Answer 1

( x <= K) 应替换为 x[1:K]。 x 是行，相应地包含 eucdistcombined/mandistcombined 的行的订单值。 ( x <= K) 仅给出 值小于 K 的索引，但是需要 最小距离值的索引 。应该是x[1:K]获取K近邻

从零开始的 k-NN 算法精度低

Low accuracy in a k-NN algorithm from scratch

r

machine-learning

knn

data-science