Huh... of course I know definition of accuracy, and how its calculated.
However even in supervised learning, accuracy is only used in very limited cases such as multi class classification. For a whole bunch of problems including the one being discussed its a poor and in some cases a biased metric. E.g. consider a heavily unbalanced problem 99% positive labels. By predicting all instances with majority label its possible to get 99% accuracy. There are several better metrics, False Accept rates, Precision Recall curves etc.
Without knowing how the dataset was collected, did the "username" leaked into the dataset, etc. its impossible to evaluate such outlandish claims.
The whole moral and ethical debate is non-sequitur, and harms legitimate deep learning research.
This entire post and the debate surrounding it, is frankly stupid. What does 95% accuracy even means?? Consider face recognition, even when there is good gold standard for matching faces (human judgement, since human are good at recognizing faces), determining accuracy of Face recognition algorithms is still challenging (E.g. Megaface challenge). When it comes to a piece of text written by an author its even more difficult. There are several practical problems too, such as how do you distinguish Quotes and copy-pasted paragraphs from rest of the text.
This sounds like a beginner who created a dataset, with a flawed metric. And is now going around claiming 95% accuracy, using "Deep" learning. And equally clueless commenters are hyping it up.
Why stop at claiming 95%, hell even I can create a "dataset" and a "deep learning" algorithm and get 99.9%.
I am not discounting that there are legitimate stylometric analysis methods, which have been peer reviewed. But please lets not hype "Deep learning for doxxing". This just sullies the real progress being made in deep learning.
However even in supervised learning, accuracy is only used in very limited cases such as multi class classification. For a whole bunch of problems including the one being discussed its a poor and in some cases a biased metric. E.g. consider a heavily unbalanced problem 99% positive labels. By predicting all instances with majority label its possible to get 99% accuracy. There are several better metrics, False Accept rates, Precision Recall curves etc.
Without knowing how the dataset was collected, did the "username" leaked into the dataset, etc. its impossible to evaluate such outlandish claims.
The whole moral and ethical debate is non-sequitur, and harms legitimate deep learning research.