I am not new to data mining, so am completely stumped with the WEKA results. Was hoping for some help. Thanks in advance!
I have a data set of numeric vectors that have a binary classification (S,H). I train a NaiveBayes model (although the method really doesn't matter) in leave one out cross-validation. The results are below:
=== Predictions on test data ===
inst# actual predicted error distribution
1 1:H 1:H *1,0
1 1:H 1:H *1,0
1 1:H 1:H *1,0
1 1:H 1:H *1,0
1 1:H 1:H *1,0
1 1:H 1:H *1,0
1 1:H 1:H *1,0
1 1:H 1:H *1,0
1 1:H 2:S + 0,*1
1 1:H 1:H *1,0
1 1:H 1:H *1,0
1 1:H 1:H *1,0
1 1:H 1:H *1,0
1 1:H 1:H *1,0
1 1:H 1:H *0.997,0.003
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 1:H + *1,0
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 2:S 0,*1
1 2:S 1:H + *1,0
=== Stratified cross-validation ===
=== Summary ===
Total Number of Instances 66
=== Confusion Matrix ===
a b <-- classified as
14 1 | a = H
2 49 | b = S
As you can see there are three errors in both the output and the confusion matrix. I then re-evaluate the model using an independent data set with the same attributes and same two classes. Here's the result:
=== Re-evaluation on test set ===
User supplied test set
Relation: FCBC_New.TagProt
Instances: unknown (yet). Reading incrementally
Attributes: 355
=== Predictions on user test set ===
inst# actual predicted error distribution
1 1:S 2:H + 0,*1
2 1:S 1:S *1,0
3 1:S 2:H + 0,*1
4 2:H 1:S + *1,0
5 2:H 2:H 0,*1
6 1:S 2:H + 0,*1
7 1:S 2:H + 0,*1
8 2:H 2:H 0,*1
9 1:S 1:S *1,0
10 1:S 2:H + 0,*1
11 1:S 2:H + 0,*1
12 2:H 1:S + *1,0
13 2:H 2:H 0,*1
14 1:S 2:H + 0,*1
15 1:S 2:H + 0,*1
16 1:S 2:H + 0,*1
17 2:H 2:H 0,*1
18 2:H 2:H 0,*1
19 1:S 2:H + 0,*1
20 1:S 2:H + 0,*1
21 1:S 2:H + 0,*1
22 1:S 1:S *1,0
23 1:S 2:H + 0,*1
24 1:S 2:H + 0,*1
25 2:H 1:S + *1,0
26 1:S 2:H + 0,*1
27 1:S 1:S *1,0
28 1:S 2:H + 0,*1
29 1:S 2:H + 0,*1
30 1:S 2:H + 0,*1
31 1:S 2:H + 0,*1
32 1:S 2:H + 0,*1
33 1:S 2:H + 0,*1
34 1:S 1:S *1,0
35 2:H 1:S + *1,0
36 1:S 2:H + 0,*1
37 1:S 1:S *1,0
38 1:S 1:S *1,0
39 2:H 1:S + *1,0
40 1:S 2:H + 0,*1
41 1:S 2:H + 0,*1
42 1:S 2:H + 0,*1
43 1:S 2:H + 0,*1
44 1:S 2:H + 0,*1
45 1:S 2:H + 0,*1
46 1:S 2:H + 0,*1
47 2:H 1:S + *1,0
48 1:S 2:H + 0,*1
49 2:H 1:S + *1,0
50 2:H 1:S + *1,0
51 1:S 2:H + 0,*1
52 1:S 2:H + 0,*1
53 2:H 1:S + *1,0
54 1:S 2:H + 0,*1
55 1:S 2:H + 0,*1
56 1:S 2:H + 0,*1
=== Summary ===
Correctly Classified Instances 44 78.5714 %
Incorrectly Classified Instances 12 21.4286 %
Kappa statistic 0.4545
Mean absolute error 0.2143
Root mean squared error 0.4629
Coverage of cases (0.95 level) 78.5714 %
Total Number of Instances 56
=== Detailed Accuracy By Class ===
TP Rate FP Rate Precision Recall F-Measure MCC ROC Area PRC Area Class
0.643 0.167 0.563 0.643 0.600 0.456 0.828 0.566 H
0.833 0.357 0.875 0.833 0.854 0.456 0.804 0.891 S
Weighted Avg. 0.786 0.310 0.797 0.786 0.790 0.456 0.810 0.810
=== Confusion Matrix ===
a b <-- classified as
9 5 | a = H
7 35 | b = S
And here is where my problems lie. The output clearly shows that there are many errors. In fact, there are 44. The confusion matrix and the result summary, on the other hand, suggest that there are 12 errors. Now, if the prediction classes were reversed, the confusion matrix would be true. So now, I look at the distribution of scores and I see that in the cross-validation results the value before the comma represents the H class, and the second value is the S class (so the value 1,0 means H prediction). However, in the test results these are reversed and the value 1,0 means S prediction. So, if I take the score distribution, the confusion matrix is right. If I take the prediction (H or S) the confusion matrix is wrong. I tried changing all test file classes to be H or S. This does NOT change the output results or the confusion matrix totals: in the confusion matrix, 16 instances are always predicted a(H) and 40 are always b(S), even though the plain text output is actually 16 b(S) and 40 a(H). Any ideas what is going wrong? It must be a simple thing, but I am completely and totally at a loss...
