I am getting a high classification accuracy(90%, for a balanced data set of size 94 ) for a two class problem with just two features. But feature analysis shows that these two features are highly correlated (98%- correlation coefficient).
When I try classification with one of these features, accuracy drops to 60%, this seems to be a puzzle for me. Any other analysis could explain this puzzle? Any help in explaining this situation is highly appreciated.
Hi Shelda,
from the examplary plot you uploaded one can see that the quantitative results you obtained are reasonable. I am not sure if there is an explanation, it could by just by chance.
Did you apply any preprocessing, e.g. apply a transform to your input data and what you show in the plot is already the transformed data? It's not common - in fact there are many transforms which do the opposite namely remove correlation between features - but in theory a transform could also introduce correlation (but i am not aware of such a transform and the goal of it).
Greetings, Sebastian
Thank you Sebastian for your response.
there was no transformation done on the data. Classification was done using actual feature values without any pre processing.
...ok, one more question: is your data 'real' data or artificial, e.g. simulated, data?
Hi Shelda,
For both of your data sets it looks like the linear fit would be quite similar (similar resulting linear equation (or y=b+ax)), yet their distribution is very different (blue class is distributed in two main clusters (top right and bottom left), whereas the orange class has no samples in the lower left section, and has a wider window for the upper right cluster (and a higher R2 overall)). Therefore, simple correlation may exist resulting from the similar linear fits, but the data distribution is quite different and, therefore, each variable provides different information to your classifying model.
Hope this helps you.
Good luck!
Hi Sebastian, I am using real data (not simulated or artificial) in this case.
Thank you John and Sebastian for your responses.
Dear Shelda,
Since your data has high correlation along the same axis, you should look for yet another set of features that are less correlated with these data.
You may need to increase the dimensionality of your problem space.
Good luck
Peyman
I too had the exact same issue. Couldn't identify what the issue was. Any help would be appreciated
- Badges
- Science topic