Open WangXuhongCN opened 5 years ago
Moreover, most algorithms use contamination proportions to determine the threshold, but I don't think GAAL is suitable for it. Since the Discriminator is a two-class classifier, the threshold should always be set as 0.5. I hope the author can reconsider that.
Thanks for the note. These two models are implemented by another contributor--I need some time to understand the underlying logic. I think that we asked for permission from the authors of the paper and have their check as well. Even though, some bugs may be there.
If you are sure these problems are legit, feel free to submit a fix or we could do something then. Thanks
In the original paper "Generative adversarial active learning for unsupervised outlier detection", the outlier score is defined as OS(x)=1-D(x) (Page 7 Algorithm 1). However, in pyod's implementation, the outlier score is defined as D(x), so I hope this mistake can be revised。