Hi thanks for the great package (and example notebooks!). My issue is summarised in two points:
It appears that feature scale influences the orientation of the hyperplane splits in the trees, resulting in a poor anomaly score map.
Is this expected behaviour? If so, can anyone offer an explanation as to how this comes about as it seems from the paper that the orientation of all hyperplanes are random.
The following illustrates this further:
I have noticed that the extended forest shows odd results when applied to features with very different scales. For example if I draw 2D points from 2 normal distributions with variance 1 and 1000 and plot the contour maps comparing the regular iForest and the extended we see the contours become horizontal and the heat map in general is not good compared to the regular iForest.
It seems as though the choice of hyperplane gets biased towards horizontal lines. This is also notable in the examples given in the paper (figure 9) where 3 plots of tree splits are shown:
Here we see the first two examples (a and b) the x and y values of the data lie on the same scale and the splits look randomly orientated. However in c) the x scale of the data is much larger than y scale, and most splits look more vertical. As a result we seen areas of higher anomaly score above and below the point cloud in the resulting heat map:
This issue is easily fixed by simply scaling all features before using the forest. However I was wondering if the splits are done on a hyperplane of random orientation why/how does feature scale influence the orientation of splits in each tree?
Apologies if I am missing something obvious, any insight would be useful, thanks!
Hi thanks for the great package (and example notebooks!). My issue is summarised in two points:
The following illustrates this further:
I have noticed that the extended forest shows odd results when applied to features with very different scales. For example if I draw 2D points from 2 normal distributions with variance 1 and 1000 and plot the contour maps comparing the regular iForest and the extended we see the contours become horizontal and the heat map in general is not good compared to the regular iForest.
It seems as though the choice of hyperplane gets biased towards horizontal lines. This is also notable in the examples given in the paper (figure 9) where 3 plots of tree splits are shown: Here we see the first two examples (a and b) the x and y values of the data lie on the same scale and the splits look randomly orientated. However in c) the x scale of the data is much larger than y scale, and most splits look more vertical. As a result we seen areas of higher anomaly score above and below the point cloud in the resulting heat map:
This issue is easily fixed by simply scaling all features before using the forest. However I was wondering if the splits are done on a hyperplane of random orientation why/how does feature scale influence the orientation of splits in each tree?
Apologies if I am missing something obvious, any insight would be useful, thanks!