wt_grnn.R

data(Boston, package = "MASS")

### CONSTRUCT THE UNWEIGHTED DATA.FRAME WITH DUPLICATES
df1 <- rbind(Boston[rep(seq(100), 5), ], Boston)
nrow(df1)
# 1006
X1 <- scale(df1[, 1:13])
Y1 <- df1[, 14]
N1 <- grnn.fit(X1, Y1)

### CONSTRUCT THE WEIGHTED DATA.FRAME WITHOUT DUPLICATES
XY <- data.frame(X1, Y1)
df2 <- Reduce(rbind, lapply(split(XY, XY[, colnames(XY)], drop = T), 
                     function(x_) data.frame(x_[1, ], cnt = nrow(x_))))
nrow(df2)
# 506
sum(df2$cnt)
# 1006
X2 <- as.matrix(df2[, 1:13])
Y2 <- df2[, 14]
W2 <- df2[, 15]
N2 <- grnn.fit(X2, Y2, W2)

### IDENTICAL PREDICTED VALUES WITH UNWEIGHTED AND WEIGHTED DATA.FRAMES
grnn.predone(N1, X1[1, ])
# 24.69219
grnn.predone(N2, X1[1, ])
# 24.69219

all.equal(grnn.predict(N1, X1[1:100, ]), grnn.predict(N2, X1[1:100, ]))
# TRUE

### COMPUTING TIME ROUGHLY LINEAR WITH RESPECT TO SIZE OF UNIQUE TRAINING SAMPLE
rbenchmark::benchmark(replications = 10, order = "elapsed", relative = "elapsed",
                      columns = c("test", "replications", "elapsed", "relative"),
  " NO WEIGHT" = grnn.predict(N1, X1[1:100, ]),
  "USE WEIGHT" = grnn.predict(N2, X1[1:100, ])
)
#         test replications elapsed relative
# 2 USE WEIGHT           10   2.157    1.000
# 1  NO WEIGHT           10   5.506    2.553