yzhao062
commented
6 years ago Thanks for asking. To my knowledge there is no universal way to generate feature importance for outlier detection. If you have the ground truth and could treat the problem as a supervised problem with imbalanced data, then you could access the feature importance with tree ensembles, e.g., xgboost.
However, if we are discussing unsupervised algorithms currently implemented in PyOD, no feature selection is available. The key reason is the result could be questionable with imbalanced data under unsupervised setting. Some outliers are well hidden in the dimensions which look unimportant.
However, this functionality might be implemented in multiple models. For instance, Isolation Forest may use the average position of the attributes in the isolation trees as the feature importance ranking. I will add this feature once the key models are implemented:)
Meanwhile, some recent researches on this topic might be helpful, like https://www.ijcai.org/Proceedings/16/Papers/272.pdf
rquintino
nehadave78
ChiragSoni95
Hello, is there anything available to identify/highlight what may be the features most probable to be triggering the outlier status? thx!