R help - Aug 2015 - kknn::predict and kknn$fitted.values

I am noticing that there is a difference between the fitted.values returned
by train.kknn, and the values returned using predict with the same model and
dataset. For example:
> data (glass)
> tmp <- train.kknn(Type ~ ., glass, kmax=1,
kernel="rectangular",
> distance=1)
> tmp$fitted.values
[[1]]
  [1] 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 1 1 1 1 1
 [62] 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 1 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2
2 1 2 1 2 2 1 2 2 5 2 2 2 6 2 2 2 2 2 2 2 2 2 2 2 2
[123] 2 2 2 2 3 2 2 2 5 5 2 2 2 2 2 2 2 2 2 2 2 2 2 2 3 3 3 3 3 3 3 3 2 3 3
3 3 3 3 2 3 7 5 5 5 5 5 5 5 5 5 5 2 5 6 6 6 6 6 6 6
[184] 2 6 7 7 2 6 7 7 7 7 7 7 7 7 7 7 7 7 5 7 7 7 7 7 7 7 7 7 7 7 7
attr(,"kernel")
[1] rectangular
attr(,"k")
[1] 1
Levels: 1 2 3 5 6 7
> predict (tmp,glass)
  [1] 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
 [62] 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2
2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2
[123] 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 3 3 3 3 3 3 3 3 3 3 3
3 3 3 3 3 3 5 5 5 5 5 5 5 5 5 5 5 5 5 6 6 6 6 6 6 6
[184] 6 6 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7
Levels: 1 2 3 5 6 7

When I check the confusion matricies for these I see that fitted.values is
giving some confusion, that is, like it is a true fit, whereas predict is
returning the exact answers.
> table (tmp$fitted.values[[1]],glass$Type)
     1  2  3  5  6  7
  1 69  4  0  0  0  0
  2  1 67  2  1  1  1
  3  0  1 15  0  0  0
  5  0  3  0 11  0  1
  6  0  1  0  0  8  1
  7  0  0  0  1  0 26
> table (predict(tmp,glass),glass$Type)   
     1  2  3  5  6  7
  1 70  0  0  0  0  0
  2  0 76  0  0  0  0
  3  0  0 17  0  0  0
  5  0  0  0 13  0  0
  6  0  0  0  0  9  0
  7  0  0  0  0  0 29

Can anyone clarify what fitted.values and predict actually do? I would have
expected they would give the same output.

Thanks... Jonathan



--
View this message in context:
http://r.789695.n4.nabble.com/kknn-predict-and-kknn-fitted-values-tp4711625.html
Sent from the R help mailing list archive at Nabble.com.

In thinking about this 'problem' last night, I found the
'solution'. Any NN
algorithm needs to keep track of all the data it is given, both X and Y
data, otherwise how could it find and report the nearest neighbour! When
predicting (i.e. predict.kknn) it will find the closest match (nearest
neighbour), which, for a point from the original dataset /is that point/!

In contrast, the kknn$fitted.values are derived from some cross validation
approach; likely either finding the nearest point with non-zero distance, or
build a model without that point and see where it falls. Otherwise, it
wouldn't be possible to report the accuracy of the model using only a single
dataset.

I will retest the algorithm using a split training/test dataset to better
understand how predict.kknn selects a model from the suite generated by
train.kknn?my original question. I assume it chooses kknn$best.parameters,
but want to verify this.

Hopefully that clarifies the issue. I post here in case future users have a
similar question. 

Thanks to any who took the time to think about this!
Jonathan



--
View this message in context:
http://r.789695.n4.nabble.com/kknn-predict-and-kknn-fitted-values-tp4711625p4711634.html
Sent from the R help mailing list archive at Nabble.com.