JSON document clustering based on schema embeddings

Abstract

The growing popularity of JSON as the data storage and interchange format increases the availability of massive multi-structured data collections. Clustering JSON documents has become a significant issue in organising large data collections. Existing research uses various structural similarity measures to perform clustering. However, differently annotated JSON structures may also encode semantic relatedness, necessitating the use of both syntactic and semantic properties of heterogeneous JSON schemas. Using the SchemaEmbed model, this paper proposes an embedding-based clustering approach for grouping contextually similar JSON documents. The SchemaEmbed model is designed using the pre-trained Word2Vec model and a deep autoencoder that considers both syntactic and semantic information of JSON schemas for clustering the documents. The Word2Vec model learns the attribute embeddings, and a deep autoencoder is designed to generate context-aware schema embeddings. Finally, the context-based similar JSON documents are grouped using a clustering algorithm. The effectiveness of the proposed work is evaluated using both real and synthetic datasets. The results and findings show that the proposed approach improves clustering quality significantly, with a high NMI score of 75%. In addition, we demonstrate that clustering results obtained by contextual similarity are superior to those obtained by traditional semantic similarity models.