Entity Consolidation: The Golden Record Problem

Four key subprocesses in data integration are: data preparation (i.e., transforming and cleaning data), schema integration (i.e., lining up like attributes), entity resolution (i.e., finding clusters of records that represent the same entity) and entity consolidation (i.e., merging each cluster into a “golden record” which contains the canonical values for each attribute). In real scenarios, the output of entity resolution typically contains multiple data formats and different abbreviations for cell values, in addition to the omnipresent problem of missing data. These issues make entity consolidation challenging. In this paper, we study the entity consolidation problem. Truth discovery systems can be used to solve this problem. They usually employ simplistic heuristics such as majority consensus (MC) or source authority to determine the golden record. However, these techniques are not capable of recognizing simple data variation, such as Jeff ↔ Jeffery, and may give biased results. To address this issue, we propose to first reduce attribute variation by merging duplicate values before applying the truth discovery system to create the golden records. For this purpose, we first align the attribute values within the same cluster to generate candidate matchings (substring pairs could be replaced by each other, e.g., 9th↔ 9 and Jeff↔ Jeffery). Then we aggregate candidate matchings with common characteristics into groups. Finally, we solicit a human to validate these matching groups and apply the approved ones to merge duplicate values. Comparing to the existing data transformation solutions, which typically try to transform an entire column from one format to another, our approach is more robust to data variety as we leverage the hidden matchings within the clusters. We tried our methods on three real world datasets. In the best case, our methods reduced the variation in clusters by 75% with high precision (>98%) by having a human confirm only 100 generated matching groups. When we invoked our algorithm prior to running MC, we were able to improve the precision of golden record creation by 40%. This work is licensed under the Creative Commons AttributionNonCommercial-NoDerivatives 4.0 International License. To view a copy of this license, visit http://creativecommons.org/licenses/by-nc-nd/4.0/. For any use beyond those covered by this license, obtain permission by emailing [email protected]. Proceedings of the VLDB Endowment, Vol. 10, No. 6 Copyright 2017 VLDB Endowment 2150-8097/17/02.