Stochastic triplet embedding

摘要: This paper considers the problem of learning an embedding data based on similarity triplets form “A is more similar to B than C”. setting relevance scenarios in which we wish model human judgements objects. We argue that order obtain a truthful underlying data, it insufficient for satisfy constraints encoded by triplets. In particular, introduce new technique called t-Distributed Stochastic Triplet Embedding (t-STE) collapses points and repels dissimilar — even when all triplet are satisfied. Our experimental evaluation three sets shows as result, t-STE much better existing techniques at revealing structure.