Accuracy issue when logically splitting dataset

[rajnepali]

Hello Friends,
I am running into an issue where I am finding it hard to explain a phenomena. Any help is much appreciated

Basically, I he a dataset with 25 attributes which is logically separated into three categories (A,B,C). When I run the test with Random Forest wit all attributes combined, I get accuracy of 93% (93/100 is correctly predicted). But when I run the test with different logical features separately and combine the results, it ends up correctly predicting more than 93 instances. To be more clear, I am attaching a picture.

As shown in the picture, combination accuracy(93)is lower than hen splitting. My best guess I that I happens because of the correlation between the features. But I am myself not satisfied with that rationale.

Thanks.

[Mark]

Try increasing the number of trees learned from the default of 100. Try 500 or 1000. You can also fiddle with the -K option - this controls the number of attributes randomly considered for splitting at a node in a given tree.

Cheers,
Mark.