Using Embeddings to Predict Changes in Large Semantic Graphs

Abstract

Understanding and predicting how large knowledge graphs change over time is as difficult as it is useful. An important subtask to address this artificial intelligence challenge is to characterize and predict three types of nodes: add-only nodes that can solely add up new edges, constant nodes whose edges remain unchanged, and del-only nodes whose edges can only be deleted. In this work, we improve previous prediction approaches by using word embeddings from NLP to identify the nodes of the large semantic graph and build a Logistic Regression model. We tested the proposed model in different versions of DBpedia and obtained the following prediction improvements on F1 measure: up to 10% for add-only nodes, close to 15% for constant nodes, and close to 22% for del-only nodes.