Towards a temporal privacy-preserving deep learning approach to residential address matching [poster]

Abstract

Integrating data from various sources has become a crucial component of decision-making processes in several application domains. In the absence of a unique identifier, addresses can serve as quasi-identifiers in the linkage of records related to the same entity in one or more data collections. Address matching is the process of identifying pairs of records by comparing full addresses or address fields, with the goal of obtaining the best matching result in relation to a searched address. Deep learning (DL) methods have gained popularity within the field of address matching due to their ability to extract semantic and contextual information from non-standard address records with redundant or missing address elements and few literal overlaps. Transformer-based algorithms are among the most used, including pretrained language models (PTLM) such as BERT and, more recently, large language models (LLMs) like ChatGPT. Based on the current state of art, two important research gaps in the field of residential address matching need to be addressed: the use of temporal data, such as creation or update timestamps, and the adoption of privacypreserving methods combining DL algorithms with the use of encrypted or non-encrypted encoding or synthetic data.