Re: Problems trained models results

> Hi all, I'm new into Weka and have some few questions yet (most surely
> noob questions, sorry).
>
> Well, this is my problem. I have trained a model (M) with 2000
> instances(X instances) via 10 cross fold validation and RandomForest
> algorithm, using 100 trees.The results are pretty decent, 74% correct
> predicted values, and a good ROC (which is what I need, not only the %
> of corrects). This info is surely not needed but well, there it goes.
>
> I want now to predict the values of 192 instances (Y instances,totally
> independent from the X ones) using the M model. But when I reevaluate
> the model the results are pretty bad. Too low % of corrects and very
> strange confussion matrix. I'm not gonna start to print all the output,
> but I think you could understand me if i put the confusion matrixes, as
> an example:
>
> Confusion matrix for X:
>
> a b c d e f <-- classified as
>
> 526 18 7 4 11 2 | a = A
>
> 38 243 28 9 12 0 | b = B
>
> 40 32 226 13 13 6 | c = C
>
> 37 12 27 222 28 4 | d = D
>
> 32 12 11 28 237 10 | e = E
>
> 21 5 7 10 29 53 | f = F
>
> Confusion matrix for Y:
>
> a b c d e f <-- classified as
>
> 42 2 1 0 5 0 | a = A
>
> 20 6 2 2 0 0 | b = B
>
> 19 4 7 0 0 0 | c = C
>
> 15 7 0 6 2 0 | d = D
>
> 15 4 2 3 5 1 | e = E
>
> 14 0 5 2 1 0 | f = F
>
> As you see, a lot of the Y instances are being predicted as "A" classes.
> I don't really have an explanation with this, and I'll be pleased with
> any info you could give me.