Detecting Media Clones in Cultural Repositories Using a Positive Unlabeled Learning Approach

Digital repositories amplify cultural-heritage research but also compound curatorial burdens such as cross-record duplicates and near-duplicates of the same artefact. We formulate curator-in-the-loop clone detection in AtticPOT as a Positive–Unlabeled (PU) learning problem. From a single anchor per artefact, a lightweight per-query Clone Encoder is trained on augmented views and uses an interpretable latent ℓ2-norm threshold to screen the unlabeled repository. The method avoids explicit negatives, exposes a transparent operating point, and integrates naturally with de-duplication, record linkage, and “find-similar” workflows.