Abstract
Classifying large sets of unlabeled entities in high-dimensional embedding spaces from a small number of labeled examples can present computational challenges, for example, when performing exhaustive nearest neighbor searches. Systems and methods can address these challenges by combining partitioned locality-sensitive hashing (LSH) with probabilistic data structures, such as multiset sketches. High-dimensional entity embeddings can be converted to partitioned LSH hashes, which may be used to build an index where each hash partition maps to a sketch aggregating information about labeled entities. For an unlabeled entity, corresponding sketches can be retrieved and merged. This final sketch can be queried to generate features based on the count of labeled neighbors that share a number of hash partitions. These neighborhood-count features may then be used as input for machine learning models or rule-based systems to facilitate scalable classification.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Golubitsky, Oleg; He, Dake; Strokach, Alexey; Berlinkov, Misha; and Abbasi, Kianoosh, "Neighborhood Inference Using Partitioned LSH and Probabilistic Sketches", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12204