Breaking the Closed-World Assumption in Stylometric Authorship Attribution

Flow of the Classify-Verify algorithm on a test document D and a suspect set A, with optional acceptance threshold t and in-set prob. p. The Classify-Verify Algorithm Input: Document D, suspect author set A = {A1, ...,An}, target measure to maximize μ Optional: in-set prob. p, manual threshold t Output: AD if AD ∈ A, and ⊥ otherwise CA← classifier trained on A VA = {VA1, ...,VAn} ← verifiers trained on A if t ,p not set then t ← threshold maximizing p-μR of ClassifyVerify cross-validation on A else if t not set then t ← threshold maximizing p-μ of ClassifyVerify cross-validation on A end if A← CA(D) if VA(D, t) = True then return A else return ⊥ end if