Combining Node Embeddings with Domain Knowledge for Identity Resolution

Abstract

The application of powerful and popular machine learning methods on real life historical data generates sub-symbolic models of the data. Such models do not perform well when trained with insufficient (or no) ground truth. We argue that the performance of these models could be improved by incorporating domain-specific knowledge, and propose an approach to incorporate symbolic domain-specific knowledge in the sub-symbolic models of the data. We show with experimental results on real historical data that our approach improves performance.