Similarity Grouping in Big Data Systems

Distributed computing technologies have opened the door for a wide range of organizations to analyze massive amounts of data. Grouping (fast but based on exact semantics) and clustering (relatively slow but based on similarity-aware semantics) are among the most useful data analysis operations. Previous work introduced the Similarity Grouping (SG) operator, which aims to integrate the best features of grouping and clustering, i.e., fast execution times and similarity-aware grouping semantics. The SG operators, however, were proposed for single node relational database systems. This paper introduces the Distributed Similarity Grouping (DSG) operator, a highly parallel operator for identifying similarity groups in big datasets. DSG enables the identification of groups where all the elements are within a given threshold from each other. This paper presents DSG's design details, implementation guidelines on Spark and Hadoop (two important Big Data systems), and extensive performance and scalability evaluation.