Scholar
Michael Boratko
Google Scholar ID: YKZGpnkAAAAJ
Research Scientist, Google
Citations & Impact
Citations
915
H-index
15
i10-index
16
Publications
20
Co-authors
63
list available
Publications
Resume (English only)
Academic Achievements
- Developed 'Box Embeddings', a region-based representation learning model that compactly represents joint probability distributions with valid set-theoretic and probabilistic semantics.
- Published 10 papers improving and extending box embeddings, achieving state-of-the-art results in collaborative filtering, textual entailment, and multi-label classification.
- Formalized the probabilistic semantics of box embeddings (UAI 2021), proving valid probability distributions even with softness, outperforming baselines.
- Proved box embeddings can represent any directed graph and introduced a trainable softness variant (NeurIPS 2021), making them optimal for directed graph representation in any dimension.
- Currently developing a measure-theoretic framework for set representation learning to establish rigorous theoretical foundations.
