Learning Embeddings from Probabilistic Triplet Comparisons
Stefan Mojsilovic · Aaltodoc (Aalto University) · 2018
Learning from relative similarity comparisons has gained interest in the data science community in the past 20 years. We introduce a new way to capture relative similarity comparisons called probabilistic triplets that alleviates extreme decisions under high uncertainty, and provides finer-grained information than ordinary triplets. We describe a new method called t-SPTE that finds an embedding of objects in a Euclidean space using probabilistic triplets datasets as its input. The problem is formulated as a least squares optimization of differences between the labeled triplet probabilities and the triplet probabilities coming from the stochastic neighborhood model in the embedding space. We experimentally show that our approach improves upon previous methods, notably t-STE, needing less labeled triplets and producing higher quality embeddings.